Research Library — Korthos
Track robotics research artifacts, datasets, benchmarks, models, repositories, and technical releases curated by Korthos.
Research Library records
- Play2Perfect studies dexterous play pretraining for precise assembly
Full research details
Summary- Introduces a two-stage RL pipeline that pretrains dexterous hands through free-space play before fine-tuning on precise contact-rich assembly tasks.
- Reports better sample efficiency than training from scratch and zero-shot sim-to-real transfer on tight insertions, multi-part assembly, and screwing tasks.
Research type PaperSource play2perfect.github.ioAuthors Tyler Ga Wei Lum, Kushal Kedia, C. Karen Liu, Jeannette BohgOrganizations Stanford UniversitySurfacedSurfacedEvidence Source Level - Match-Any-Events adds zero-shot feature matching for event cameras
Full research details
Summary- Provides an event-camera feature matching method intended for wide-baseline, motion-robust perception.
- Includes a GitHub release, making the artifact actionable for perception stacks that use event cameras.
Research type Code ReleaseSource github.comAuthors Ziyun WangOrganizations Johns Hopkins UniversitySurfacedSurfacedEvidence Source Level - SwarmSubnet benchmarks simulated drone autonomy tasks
Full research details
Summary- Presents a simulated drone autonomy benchmark focused on navigation and swarm-oriented tasks.
- Adds useful coverage outside humanoids and manipulators for the research attention directory.
Research type BenchmarkSource intotao.appAuthors IntoTAOOrganizations IntoTAOSurfacedSurfacedEvidence Source LevelArtifact SwarmSubnet - ManipArena releases a real-world reasoning benchmark for generalist robot manipulation
Full research details
Summary- Standardizes real-world evaluation for reasoning-oriented generalist robot manipulation with 20 tasks and 10,812 expert trajectories.
- Targets the gap between simulator-centric benchmarks and real manipulation systems with out-of-distribution settings, mobile manipulation, and sensory diagnostics.
Research type BenchmarkSource arxiv.orgAuthors Yu Sun, Meng Cao, Ping Yang, Rongtao Xu, Yunxiao Yan, Runze Xu, Liang Ma, Roy Gan, Andy Zhai, Qingxuan Chen, Zunnan Xu, Hao Wang, Jincheng Yu, Lucy Liang, Qian Wang, Ivan Laptev, Ian D. Reid, Xiaodan LiangOrganizations Sun Yat-sen University, MBZUAISurfacedSurfacedEvidence Source Level - HIW-500 LeRobot format makes a humanoid teleoperation dataset easier to reuse
Full research details
Summary- Packages the HIW-500 humanoid teleoperation dataset in LeRobot format for easier training and reuse.
- Useful for humanoid manipulation and imitation-learning teams that need large-scale real-world teleop traces.
Research type DatasetSource huggingface.coAuthors Unitree Robotics, Hugging Face LeRobotOrganizations Unitree Robotics, Hugging FacePublished - ABot-Explorer releases an embodied exploration framework with scene-graph memory
Full research details
Summary- Open framework for online embodied exploration and memory construction using VLM-derived semantic navigational affordances.
- Builds scene-graph memory during exploration and targets RGB-only embodied navigation settings.
- Repository and paper support IROS 2026 evaluation with reported gains in efficiency and coverage.
Research type Github RepoSource githubAuthors AMAP CV Lab teamOrganizations AMAP CV LabSurfacedSurfacedEvidence Strong - Universal Manipulation Exoskeleton adds torque-feedback teleoperation for compliant robot policies
Full research details
Summary- Introduces a low-cost upper-limb exoskeleton that provides real-time haptic torque feedback while recording whole-arm configuration and joint torque signals.
- Retargets to multiple robot arms including OpenArm, Franka, and X-ARM for bimanual, mobile, and constrained-space manipulation.
- Project page and paper report improved compliant policy learning on contact-rich tasks, including force-mediated and visually occluded manipulation.
Research type Project PageSource project_pageAuthors Litian Liang, Jingxi Xu, Xinda Qi, Yujun Cai, Houzhu Ding, Luqi Wang, Zhixin Sun, Jyh-Herng Chow, Ming Yang, Mark CutkoskyOrganizations Inclusion AI, Stanford UniversitySurfacedSurfacedEvidence Strong - Humanoid Everyday focuses on daily-task humanoid manipulation
Full research details
Summary- Studies general-purpose humanoid manipulation in everyday task settings.
- Adds a clear humanoid manipulation artifact to the social-attention research set.
Research type PaperSource huggingface.coAuthors Humanoid Everyday Project TeamOrganizations Humanoid EverydayPublishedArtifact Humanoid Everyday - HumanScale shows filtered egocentric human video can outperform robot data for embodied pretraining
Full research details
Summary- Studies whether filtered egocentric human videos with pseudo-action labels can outperform matched real-robot teleoperation data for embodied pretraining.
- Reports better action-prediction validation loss and higher real-robot task success across in-distribution and out-of-distribution settings.
- Supports a scalable pretrain-on-human-video, adapt-with-robot-data paradigm for embodied robot policies.
Research type PaperSource arxivAuthors Juncheng Ma, Jianxin Bi, Yufan DengOrganizations NVIDIA, Academic collaboratorsSurfacedSurfacedEvidence Strong - UMA uses 3D object motion as an embodiment-agnostic interface for robot foundation models
Full research details
Summary- Introduces a robot foundation model that uses 3D object motion as a shared embodiment-agnostic interface between perception and action.
- Models motion and action as co-evolving variables in one generative framework for zero-shot visuomotor control, dynamics prediction, and few-shot task-token adaptation.
- Project page reports results across 6-DoF manipulation, tool use, and deformable-object tasks, with language or goal-image conditioning.
Research type Project PageSource project_pageAuthors Yunhao Cao, Shitong Liu, Chaofeng Wang, Meryl Zhang, Xuanchen LuOrganizations Cornell UniversitySurfacedSurfacedEvidence Strong - Playful Agentic Robot Learning uses robot-agent teams to build reusable skills before tasks
Full research details
Summary- Studies self-directed robot play as a continual pre-task skill-acquisition phase.
- Introduces Robotics Agent Teams that propose exploratory tasks, execute code-as-policy, verify progress, diagnose failures, and build a reusable skill library.
- Reports downstream task gains from play-learned skills and transfer to other agents without fine-tuning.
Research type PaperSource arxivAuthors Junyi Zhang, Jiaxin Ge, Hanjun YooOrganizations UC Berkeley, CollaboratorsSurfacedSurfacedEvidence StrongArtifact Playful Agentic Robot Learning - Human Universal Grasping releases human-data method and benchmark for zero-shot dexterous robot grasping
Full research details
Summary- Introduces HUG, a flow-matching model that predicts diverse human grasps from a single RGB-D image.
- Trains on 1M-HUGs, an egocentric human grasp dataset spanning more than one million frames and thousands of objects.
- Includes HUG-Bench and real-world retargeting results across robot hand embodiments.
Research type PaperSource project_siteAuthors Kevin Yuanbo Wu, Tianxing Zhou, Isaac Tu, Billy Yan, Irmak GuzeyOrganizations New York University, Tsinghua University, University of MichiganSurfacedSurfacedEvidence StrongArtifact Human Universal Grasping - UAV-CAS releases a calibrated digital-twin dataset for UAV swarm intrusion detection
Full research details
Summary- Publishes a labeled flow dataset for UAV swarm network intrusion detection generated from a calibrated Containernet digital twin.
- Uses AERPAW testbed measurements to calibrate path loss, mobility, link performance, and end-to-end trace fidelity.
- Supports drone swarm security and aerial robotics network-resilience research.
Research type DatasetSource arxivAuthors Sripath Mishra, Bharat Bhargava, Zizheng Liu, Shafkat IslamOrganizations Purdue UniversitySurfacedSurfacedEvidence Strong - Alibaba Qwen releases Qwen-Robot Suite for navigation, manipulation, and robot world modeling
Full research details
Summary- Introduces Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld as a foundation-model suite for physical-world intelligence.
- Covers language-conditioned navigation, cross-embodiment manipulation, and video world modeling for robot scenarios.
- Provides technical reports and model resources from the Qwen research program.
Research type Model ReleaseSource project_siteAuthors Qwen TeamOrganizations Alibaba, QwenSurfacedSurfacedEvidence Strong - JAXA publishes Science Robotics results from the LEV-2 transformable lunar rover
Full research details
Summary- Publishes results from LEV-2, also known as SORA-Q, a palm-sized transformable rover deployed during the SLIM lunar mission.
- Details autonomous transformation, lunar locomotion, imaging, and operation on the Moon.
- Demonstrates compact space robotics hardware in an extreme real-world environment.
Research type PaperSource official_pressAuthors JAXA team, TOMY Company, Sony Group, Doshisha UniversityOrganizations JAXA, TOMY Company, Sony Group, Doshisha UniversitySurfacedSurfacedEvidence Strong - Ai2 releases MolmoMotion-1M, model weights, and PointMotionBench for language-guided 3D motion forecasting
Full research details
Summary- Publishes MolmoMotion-1M, a dataset of 1.16 million videos with action descriptions and object-grounded 3D point trajectories.
- Releases MolmoMotion model weights and PointMotionBench for human-validated object-centric 3D motion forecasting.
- Targets downstream robotics, world modeling, egocentric video understanding, and controllable video generation.
Research type DatasetSource technical_blogAuthors AllenAI teamOrganizations Allen Institute for AISurfacedSurfacedEvidence Strong - Stringman opens a ceiling-mounted LeRobot platform for household cleanup
Full research details
Summary- Open-source overhead cable-driven parallel robot that hangs from the ceiling to avoid floor-navigation problems in cluttered homes.
- Uses LeRobot-based firmware, 3D-printed hardware, build guides, and a local multitask diffusion-transformer policy trained on a small teleoperation dataset.
- Demonstrates autonomous pickup of laundry, toys, and trash with a low-cost home-oriented robot morphology.
Research type Github RepoSource githubAuthors Nicholas HnifongOrganizations Neufangled Robotics, LeRobot communitySurfacedSurfacedEvidence Strong - Double Preconditioning targets test-time feedback mismatch in robot policy learning
Full research details
Summary- Introduces Double Preconditioning as a plug-in optimizer design for settings where validation loss diverges from deployment success.
- Targets test-time feedback mismatch in robot policy learning and other long-horizon rollout settings.
- Evaluates image-based robot policy learning on challenging Robomimic tasks.
Research type PaperSource arxivAuthors Thomas T. Zhang, Alok Shah, Yifei Zhang, Vincent Zhang, Nikolai Matni, Max SimchowitzOrganizations University of PennsylvaniaSurfacedSurfacedEvidence Strong - XDOF releases ABC-130K as an open bimanual teleoperation dataset and training stack
Full research details
Summary- Publishes 130,000-plus bimanual teleoperation trajectories and roughly 3,500 hours of manipulation data collected on low-cost dual-arm YAM stations.
- Includes Apache-2.0 data, training code, simulator assets, and evaluation resources for scalable behavior cloning.
- Covers 195 manipulation tasks with RGB, action, trajectory, and proprioceptive signals.
Research type DatasetSource huggingfaceAuthors Arthur Allshire, Himanshu Gaurav Singh, Ritvik Singh, Adam Rashid, Hongsuk Choi, David McAllister, Philipp Wu, Fred Shentu, Guanya ShiOrganizations XDOF, UC Berkeley, Carnegie Mellon University, MIT, Amazon FARSurfacedSurfacedEvidence Strong - NVIDIA, CMU, and UC Berkeley introduce ENPIRE for agentic robot policy self-improvement
Full research details
Summary- Introduces a closed-loop physical autoresearch system that lets coding agents reset tasks, run real-robot rollouts, verify outcomes, and refine policies.
- Reports 99% pass@8 success on dexterous manipulation tasks including pin insertion, zip-tie cutting, and GPU insertion.
- Shows fleet scaling across one, four, and eight agent-robot teams for faster real-world policy improvement.
Research type Project PageSource project_siteAuthors Wenli Xiao, Jia Xie, Tonghe Zhang, Haotian Lin, Letian Fu, Haoru Xue, Linxi Fan, Yuke Zhu, Guanya ShiOrganizations NVIDIA, Carnegie Mellon University, UC BerkeleySurfacedSurfacedEvidence Strong - ACE Robotics releases Kairos world model resources for embodied intelligence benchmarks
Full research details
Summary- Open repository and Hugging Face organization provide Kairos world-model resources for embodied agents.
- Positions the model around multimodal understanding, generation, action prediction, and long-horizon robot reasoning.
- June 15 coverage highlights benchmark performance across RoboTwin, LIBERO-Plus, WorldModelBench Robot, and DreamGen.
Research type Model ReleaseSource githubAuthors ACE Robotics teamOrganizations ACE RoboticsSurfacedSurfacedEvidence StrongArtifact Kairos World Model - Tencent details Hy-Embodied-0.5-VLA as a real-world VLA stack from data to deployment
Full research details
Summary- Presents an end-to-end robot-learning stack covering data collection, model design, pretraining, supervised fine-tuning, RL post-training, and deployment.
- Uses large egocentric UMI-style data and FlowPRO-style offline reinforcement learning for practical manipulation tasks.
- Reports cross-embodiment transfer and real-world task performance on tasks such as mug hanging, stacking, and cabinet operation.
Research type PaperSource arxivAuthors He Zhang, Lingzhu Xiang, Haitao Lin, Zeyu Huang, Minghui Wang, Dingyan Zhong, Yubo Dong, Yihao Wu, Yongming Rao, Dongsheng Zhang, Wanjia He, Ling Chen, Kai Huang, Jiahao Chen, Sichang Su, Xumin Yu, Ziyi Wang, Chengwei Zhu, Xiao Teng, Yuchun Guo, Yufeng Zhang, Yuandong Liu, Rui Wang, Zisheng Lu, Han Hu, Zhengyou ZhangOrganizations Tencent Robotics X, Tencent HunyuanSurfacedSurfacedEvidence Strong - Swarm releases SOTApilot as an open drone-navigation stack for a 1,000-seed UAV benchmark
Full research details
Summary- Open-source drone-navigation stack reports 95.34% success on Swarm's public 1,000-seed benchmark.
- Covers general UAV navigation across six procedural environments.
- Combines obstacle avoidance, objective recognition, and autonomous state-machine control.
Research type BenchmarkSource githubAuthors Swarm teamOrganizations SwarmSurfacedSurfacedEvidence StrongArtifact SOTApilot - DeepGlint releases FAST-LIO humanoid localization code for LiDAR-based robot navigation
Full research details
Summary- Open-sources a LiDAR localization stack adapted for humanoid platforms such as Unitree G1-style robots.
- Builds on FAST-LIO with humanoid-oriented deployment changes and ROS/C++ prerequisites.
- Adds practical perception/navigation infrastructure for stable humanoid operation in unstructured environments.
Research type Github RepoSource githubAuthors Axiang Sun, Junfeng Zhu, Yihao Xu, Xin Li, Zhongxia Zhao, Gaohao Zhou, Dongliang Li, Xiang An, Huajie Tan, Ziyong FengOrganizations DeepGlintSurfacedSurfacedEvidence Strong - NavWAM introduces a navigation world-action model for goal-conditioned visual navigation
Full research details
Summary- Jointly predicts future egocentric observations, goal progress, and action chunks for visual navigation.
- Runs as a closed-loop diffusion-transformer policy without an external planner.
- Project page reports benchmark gains and real-robot deployment on a Diablo mobile platform; code is marked coming soon.
Research type Project PageSource project_pageAuthors Daichi Azuma, Taiki Miyanishi, Koya Sakamoto, Shuhei Kurita, Yaonan Zhu, Petr Khrapchenkov, Motoaki Kawanabe, Yusuke Iwasawa, Yutaka MatsuoOrganizations University of Tokyo, RIKEN AIPSurfacedSurfacedEvidence Strong - RoboProcessBench benchmarks process-aware understanding in vision-language robotic manipulation
Full research details
Summary- Introduces a benchmark for whether VLMs and VLAs understand how manipulation processes unfold, not just final outcomes.
- Targets sequential, process-aware reasoning in vision-language robotic manipulation.
- Useful for evaluating long-horizon embodied policies where intermediate state changes matter.
Research type BenchmarkSource arxivAuthors Dayu Xia, Yue Shi, Yao Mu, Huiting Ji, Chaofan Ma, Yingjie Zhou, Hua Chen, Yang Liu, Jiezhang Cao, Guangtao ZhaiOrganizations Shanghai Jiao Tong UniversitySurfacedSurfacedEvidence Strong - Flow Reversal Steering improves generalist robot policies from coarse human or VLM guidance
Full research details
Summary- Uses flow reversal to refine coarse reference actions from humans or VLMs into precise in-distribution robot actions.
- Supports zero-shot control, fast distillation into smaller policies, and bootstrapped RL improvement.
- Project page reports results on LIBERO simulation and DROID real-world manipulation; code is marked coming soon.
Research type Project PageSource project_pageAuthors Andy Tang, William Chen, Andrew Wagenmaker, Chelsea Finn, Sergey LevineOrganizations Stanford University, UC BerkeleySurfacedSurfacedEvidence Strong - XRZero-G0 opens a robot-learning framework and multimodal G0 dataset for dexterous manipulation
Full research details
Summary- Open-sources a hardware/software framework for scalable robot learning with human-operated data collection and inspection pipelines.
- Pairs the framework with a Hugging Face G0 dataset surface for multimodal dexterous manipulation data.
- Targets reduced real-robot data requirements through interfaces, quality control, and better data ratios.
Research type Open SourceSource githubAuthors X Square Robot teamOrganizations X Square RobotSurfacedSurfacedEvidence Strong - WEAVER introduces an efficient latent world model for long-horizon robotic manipulation
Full research details
Summary- Presents an action-conditioned latent world model for future-state prediction, multi-view decoding, and latent-space planning.
- Focuses on fidelity, long-horizon consistency, and fast inference through sparse memory, KV caching, and rectified flow.
- Project page reports policy evaluation, policy improvement, and test-time planning results on DROID/OOD manipulation data.
Research type Project PageSource project_pageAuthors Arnav Kumar Jain, Yilin Wu, Jesse Farebrother, Gokul Swamy, Andrea BajcsyOrganizations Carnegie Mellon University, UC BerkeleySurfacedSurfacedEvidence Strong - World Model Self-Distillation releases code, weights, and data for task-solving video world models
Full research details
Summary- Turns pretrained caption-guided video generators into instruction-conditioned executors through self-distillation and VLM feedback.
- Releases a project page, GitHub repository, Hugging Face model weights, and the WorldTasks dataset.
- Reports transfer to robotic-task benchmarks including DreamGen-style physical task execution.
Research type Model ReleaseSource project_pageAuthors Sebastian Stapf, Pablo Acuaviva Huertos, Aram Davtyan, Paolo FavaroOrganizations University of BernSurfacedSurfacedEvidence Strong - Embodied-R1.5 releases an open embodied foundation model stack for planning and robot control
Full research details
Summary- Releases an 8B embodied foundation model spanning embodied cognition, planning, correction, and pointing.
- Combines a 15B+ token embodied dataset with multi-task reinforcement learning and a Planner-Grounder-Corrector loop for long-horizon self-correction.
- Project page reports open weights, datasets, code, evaluation tooling, benchmark results, and zero-shot real-robot transfer.
Research type Model ReleaseSource project_pageAuthors Yifu Yuan, Yaoting Huang, Xianze Yao, Yutong Li, Shuoheng Zhang, Linqi Han, Pengyi Li, Jiangeng Sun, Wenting Jia, Zhao Zhang, Yuhao Liu, Ruihao Liao, Yucheng Hu, Qiyu Wu, Yuxiao Li, Zibin Dong, Fei Ni, Yan Zheng, Shuyang Gu, Yi Ma, Hongyao Tang, Han Hu, Jianye HaoOrganizations Tianjin University, Tencent HunyuanSurfacedSurfacedEvidence Strong - Mana shows zero-shot sim-to-real dexterous manipulation of articulated tools
Full research details
Summary- Frames dexterous articulated-tool manipulation as an animation-style pipeline for tools such as tongs, pliers, clothespins, and syringes.
- Combines procedural grasp keyframes, motion planning, and reinforcement learning for trajectories.
- Project page reports zero-shot sim-to-real transfer with minimal per-tool user input and foundation-model point-cloud observations.
Research type Project PageSource project_pageAuthors Zhao-Heng Yin, Guanya Shi, Pieter Abbeel, C. Karen LiuOrganizations UC Berkeley, CMU, Stanford University, Amazon FARSurfacedSurfacedEvidence Strong - Robots Need More than VLA and World Models frames missing interfaces for robot learning
Full research details
Summary- Argues that robot policy scaling is bottlenecked by converting unstructured data into grounded supervision.
- Frames four missing interfaces: behavior autolabelling, embodiment retargeting, physics-grounded 3D world modeling, and reward inference.
- High-engagement position paper gives useful strategic context for the VLA and world-model wave.
Research type PaperSource arxivAuthors Elis Karcini, Faisal Mehrban, Quang Nguyen, Mac Schwager, Arash Ajoudani, Cesar Cadena, Jan Peters, Marco Hutter, Haitham Bou-AmmarOrganizations Stanford University, ETH Zurich, TU DarmstadtSurfacedSurfacedEvidence Strong - MotionDisco project released for search-discovered humanoid loco-manipulation skills
Full research details
Summary- Introduces an automated framework for discovering long-horizon, contact-rich humanoid loco-manipulation skills without human motion capture, teleoperation, or retargeting.
- Combines LLM-guided evolutionary search over contact plans, kinodynamic optimization, pruning, and reinforcement-learning tracking policies.
- Project page and arXiv paper report transfer of discovered trajectories to real humanoid hardware across cluttered scenes and multi-object transport tasks.
Research type Project PageSource project_pageAuthors Ilyass Taouil, Michal Ciebielski, Shafeef Omar, Haizhou Zhao, Angela Dai, Aaron M. Johnson, Majid KhadivOrganizations Technical University of Munich, New York University, Carnegie Mellon UniversitySurfacedSurfacedEvidence Strong - Dexora releases open high-DoF bimanual VLA system for dexterous manipulation
Full research details
Summary- Releases a high-DoF bimanual VLA system targeting dual-arm, dual-hand dexterous manipulation.
- Project page links arXiv, code, and dataset resources and reports 100k simulated trajectories plus 10k real teleoperated episodes.
- Demonstrates cross-embodiment transfer from a 36-DoF dual-arm dual-hand setup to lower-DoF gripper and arm-hand embodiments.
Research type Project PageSource project_pageAuthors Zongzheng Zhang, Jingrui Pang, Zhuo Yang, Kun Li, Minwen Liao, Saining Zhang, Guoxuan Chi, Jinbang Guo, Huan-ang Gao, Modi Shi, Dongyun Ge, Yao Mu, Jiayuan Gu, Rui Chen, Hao Dong, Huazhe Xu, Li Yi, Yixin Zhu, Hang Zhao, Pengwei Wang, Shanghang Zhang, Guocai Yao, Jianyu Chen, Hongyang Li, Hao ZhaoOrganizations AIR-SUN, Tsinghua University, OpenDriveLabSurfacedSurfacedEvidence Strong - NVIDIA releases Anchor Lab dataset for real-world sim-to-real calibration
Full research details
Summary- Releases a Hugging Face robotics dataset for grounding simulation calibration in real physical measurements.
- Supports zero-shot sim-to-real workflows by providing hardware measurement data for dynamics and calibration pipelines.
- Apache-2.0 dataset surface is live on Hugging Face and is part of NVIDIA physical-AI infrastructure.
Research type DatasetSource huggingfaceAuthors NVIDIA Robotics teamOrganizations NVIDIASurfacedSurfacedEvidence Strong - NVIDIA releases Cosmos 3 omnimodal world models for physical AI
Full research details
Summary- Releases the Cosmos 3 model collection for omnimodal world modeling across language, vision, video, audio, and robot-action settings.
- Includes open model weights plus NVIDIA Cosmos code and project resources for physical-AI generation, reasoning, and policy use cases.
- The Hugging Face collection includes Cosmos3 model variants, policy-oriented models, and synthetic physical-AI datasets.
Research type Model ReleaseSource huggingfaceAuthors NVIDIA Cosmos Lab teamOrganizations NVIDIASurfacedSurfacedEvidence Strong - RoboSemanticBench releases open benchmark for VLA semantic grounding
Full research details
Summary- Provides an embodied benchmark for diagnosing semantic grounding in VLA action prediction.
- Uses multiple-choice math and general-knowledge tasks where the robot must select and manipulate the physically correct answer block.
- Repository includes benchmark suites, RoboTwin 2.0 integration, data-collection workflows, and policy-evaluation scripts.
Research type RepoSource githubAuthors Bin Yu, Yao Zhang, Haishan Liu, Shijie Lian, Yuliang Wei, Xiaopeng Lin, Zhaolong Shen, Changti Wu, Ruina Hu, Bailing Wang, Cong Huang, Kai ChenOrganizations ZGC-EmbodyAISurfacedSurfacedEvidence Strong - LEGS project page released for teleop-free humanoid VLA fine-tuning
Full research details
Summary- Introduces a teleoperation-free VLA fine-tuning pipeline for humanoid loco-manipulation in an embodied Gaussian Splatting world.
- Uses MuJoCo physics, photorealistic 3DGS backgrounds, SAM3D objects, and motion primitives to generate labeled demos.
- Reports Unitree G1 pick-and-place results with re-rendered data enabling cheaper robustness training.
Research type Project PageSource project_pageAuthors Hojune Kim, Timothy Chen, Jiankai Sun, Lars W. Osterberg, Qianzhong Chen, Ke Wang, Mac SchwagerOrganizations Stanford UniversitySurfacedSurfacedEvidence Strong - ETH releases WholeBodyLocomotion project for terrain-aware humanoid control
Full research details
Summary- Combines online terrain-aware motion generation with reinforcement-learning motion tracking for whole-body humanoid locomotion.
- Deploys on a Unitree G1 with onboard perception across boxes, hurdles, stairs, and mixed terrain combinations.
- The project page provides paper and video resources; no public code link was verified on the page during this pass.
Research type Project PageSource project_pageAuthors Zewei Zhang, Kehan Wen, Michael Xu, Junzhe He, Chenhao Li, Takahiro Miki, Clemens Schwarke, Chong Zhang, Xue Bin Peng, Marco HutterOrganizations Robotic Systems Lab, ETH Zurich, EPFL, Simon Fraser University, ETH AI Center, NVIDIASurfacedSurfacedEvidence Strong - Humanoid-GPT releases code for billion-frame zero-shot humanoid motion tracking
Full research details
Summary- Releases a GPT-style humanoid motion Transformer trained on a 2B-frame retargeted motion corpus for zero-shot motion tracking.
- Repository includes inference, evaluation, MuJoCo/MJX simulation, released checkpoints, and Unitree G1 deployment paths.
- Project page reports real-time whole-body control and zero-shot tracking on Unitree G1 across everyday home and high-dynamic motion scenarios.
Research type RepoSource githubAuthors Zekun Qi, Xuchuan Chen, Dairu Liu, Chenghuai Lin, Yunrui Lian, Sikai Liang, Zhikai Zhang, Yu Guan, Jilong Wang, Wenyao Zhang, Xinqiang Yu, He Wang, Li YiOrganizations Tsinghua University, Galbot Inc., Beihang University, Shanghai Jiao Tong University, Peking University, Shanghai Qi Zhi InstituteSurfacedSurfacedEvidence Strong - NVIDIA announces Isaac GR00T open reference humanoid for academic research
Full research details
Summary- Announces an open humanoid reference design combining a Unitree H2 Plus body, Sharpa tactile hands, Jetson Thor onboard compute, and Isaac GR00T software workflows.
- Targets academic humanoid research with a unified path for data capture, simulation, policy learning, evaluation, and real-world deployment.
- Hardware is slated for late 2026 and the Unitree G1 workflow is listed as coming soon on GitHub and Hugging Face.
Research type Project PageSource nvidiaAuthors NVIDIA Robotics teamOrganizations NVIDIA, Unitree, Sharpa, Ai2, ETH Zurich, Stanford University, UC San DiegoSurfacedSurfacedEvidence Strong - XPENG Robotics publishes Fe0 cross-embodiment foundation-model study
Full research details
Summary- Analyzes what large cross-embodiment data teaches an embodied foundation model under minimal in-domain humanoid anchors.
- Evaluates transfer from heterogeneous robot and human video data onto the XPENG IRON-R01 humanoid.
- Project page is live, but no public code or weights were verified in this pass.
Research type Project PageSource project_pageAuthors Yizhuo Li, Xiaoyu Zhang, Yuying Ge, Teng Wang, Boyu Chen, Yi Chen, Bo Liu, Feng Qiu, Yuguo Gan, Hui Zhou, Yixiao GeOrganizations XPENG RoboticsSurfacedSurfacedEvidence Strong - MIT releases SOLE-R1 open model for zero-shot reward prediction in online robot RL
Full research details
Summary- Releases an open video-language reasoning model that produces per-timestep chain-of-thought traces and dense task-progress rewards from video observations and natural-language goals.
- Includes code, evaluation examples, model outputs, the SOLE-R1-8B checkpoint, and a large Hugging Face training dataset for reward reasoning over robot videos.
- Shows zero-shot online reinforcement learning from scratch across simulated manipulation suites and real Franka tabletop tasks without demonstrations, ground-truth rewards, or task-specific tuning.
Research type RepoSource githubAuthors Philip Schroeder, Thomas Weng, Karl Schmeckpeper, Eric Rosen, Stephen Hart, Ondrej BizaOrganizations MIT CSAIL, MIT EECSSurfacedSurfacedEvidence Strong - Real2Sim-Eval releases Gaussian-splatting digital twins for robot policy evaluation
Full research details
Summary- Builds real-to-sim soft-body digital twins from real-world videos using Gaussian Splatting for rendering and optimized physical parameters for dynamics.
- Evaluates policies in closed loop on toy packing, rope routing, and T-block pushing, with simulated outcomes correlating with real-world rollouts.
- Project page links paper, code, and a Hugging Face collection containing datasets, policies, PhysTwin assets, and Gaussian Splatting checkpoints.
Research type Project PageSource project_pageAuthors Kaifeng Zhang, Shuo Sha, Hanxiao Jiang, Matthew Loper, Hyunjong Song, Guangyan Cai, Zhuo Xu, Xiaochen Hu, Changxi Zheng, Yunzhu LiOrganizations Columbia University, SceniX Inc., Google DeepMindSurfacedSurfacedEvidence Strong - RoboMIND V2 releases 107k-trajectory multi-embodiment manipulation dataset and benchmark
Full research details
Summary- Expands RoboMIND V2 to 107k real-world teleoperation trajectories across 479 tasks, 96 object classes, and four robot embodiments.
- Includes multi-view RGB, proprioception, language annotations, failure cases, and an Isaac Sim digital twin for synthetic data and evaluation.
- Provides public project, data, Hugging Face, GitHub, and paper links for imitation-learning and VLA training benchmarks.
Research type DatasetSource project_pageAuthors RoboMIND TeamOrganizations X-Humanoid, Beijing Innovation Center of Humanoid Robotics, Peking University, Beijing Academy of Artificial IntelligenceSurfacedSurfacedEvidence Strong - AGIBOT releases tau0-WM open code and VAM weights for video-action world modeling
Full research details
Summary- Releases the official implementation for a unified 5B video-action world model for robotic manipulation, alongside public VAM model weights on Hugging Face.
- Trains on roughly 27.3k hours of heterogeneous embodied data spanning real-robot teleoperation, UMI-style interaction data, egocentric human video, and rollout or failure trajectories.
- Supports real-world deployment through a policy server and action-generation pipeline; simulator weights and test-time-computation code are still marked coming soon.
Research type RepoSource githubAuthors Pengfei Zhou, Shengcong Chen, Di Chen, Jiaxu Wang, Rongjun Jin, Bingwen Zhu, Yike Pan, Songen Gu, Kuanning Wang, Shufeng Nan, Xingyu Qiu, Chenhao Qiu, Pu Yang, Yunuo Cai, Jianxiong Gao, Yifan Li, Yanwei Fu, Xiangyu Yue, Zhi Chen, Jianlan LuoOrganizations AGIBOT Finch, Shanghai Innovation InstituteSurfacedSurfacedEvidence Strong - PhAIL releases real-robot VLA benchmark with distributional evaluation methodology
Full research details
Summary- Launches a real-robot VLA benchmark on Franka FR3 with distributional metrics beyond binary success rates.
- Includes analysis code, paper sources, annotation pipeline, dataset metadata, and public dataset download instructions.
- Reports throughput-style comparisons across public VLA systems using time-to-success curves, bootstrap confidence intervals, and KS tests.
Research type BenchmarkSource phail.aiAuthors Sergey ArkhangelskiyOrganizations Positronic RoboticsSurfacedSurfacedEvidence Strong - Wall-OSS-0.5 releases open VLA model weights for zero-shot real-robot manipulation
Full research details
Summary- Releases Wall-OSS-0.5 model weights and Wall-X infrastructure for embodied foundation model training and inference.
- Targets zero-shot real-robot manipulation with flow-matching and FAST action branches plus LeRobot-compatible data preparation.
- Repository provides training/inference and evaluation utilities; the README still marks Wall-OSS-0.5-specific code as coming soon.
Research type Model ReleaseSource githubAuthors Andy Zhai, Brae Liu, Bruno Fang, Chalse Cai, Ellie Ma, Ethan Yin, Hao Wang, Hugo Zhou, James Wang, Lights Shi, Lucy Liang, Make Wang, Qian Wang, Roy Gan, Ryan Yu, Shalfun Li, Starrick Liu, Sylas Chen, Vincent Chen, Zach XuOrganizations X-Square-RobotSurfacedSurfacedEvidence Strong - Toyota releases YUBI open hardware for finger-driven teleoperation and dexterous manipulation
Full research details
Summary- Releases open hardware for a wearable YUBI Glove and robot-mounted YUBI Gripper for finger-driven teleoperation.
- Provides CAD, STL print files, BOM CSVs, assembly PDFs, media assets, and a companion software repository for control and teleoperation.
- Designed as a replicable hardware interface for dexterous manipulation data collection and robot-learning pipelines.
Research type RepoSource githubAuthors Yoshihiro Okumatsu, Daiki Fukunaga, Jumpei Arima, Yuki NoguchiOrganizations Frontier Research Center, Toyota Motor CorporationSurfacedSurfacedEvidence Strong - MIDAS Hand releases open modular dexterous sensing hand hardware and software
Full research details
Summary- Releases a modular low-impedance direct-drive anthropomorphic hand intended for dexterous manipulation research.
- Provides open hardware and software components including API, retargeting, MuJoCo simulation, teleoperation, and communication-board repositories.
- Designed for integration with robot arms, teleoperation systems, and robot-learning data collection pipelines.
Research type RepoSource githubAuthors MIDAS Hand contributorsOrganizations MIDAS Hand projectSurfacedSurfacedEvidence Strong - Qwen-VLA repo released for unified generalist VLA across tasks and robot embodiments
Full research details
Summary- Introduces a unified VLA generalist using a Qwen3.5-4B vision-language backbone and DiT flow-matching action decoder.
- Casts manipulation, navigation, and trajectory prediction into one action-and-trajectory framework with embodiment-aware prompt conditioning.
- Reports simulation and real-world ALOHA bimanual results; the repository includes report, blog, benchmarks, and demo but no explicit model weights release yet.
Research type RepoSource githubAuthors Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, Xuhong Huang, Pei Lin, Junyang Lin, Dayiheng Liu, Shuai Bai, Jingren Zhou, Jiazhao Zhang, Haoqi Yuan, Gengze Zhou, Hang Yin, Ye Wang, Yiyang Huang, Zixing Lei, Wujian Peng, Delin Chen, Yingming Zheng, Jingyang Fan, Xianwei Zhuang, Xin Zhou, Haoyang Li, Anzhe Chen, Tong Zhang, Xuejing Liu, Yuchong Sun, Ruizhe Chen, Zhaohai Li, Chenxu Lu, Zhibo Yang, Tao Yu, Xionghui ChenOrganizations Qwen, Alibaba Tongyi LabSurfacedSurfacedEvidence Strong - DynaFLIP project page released for tri-modal dynamics-guided manipulation representations
Full research details
Summary- Introduces a dynamics-aware multimodal pre-training framework that aligns image transitions, language, and 3D flow to produce manipulation-relevant visual representations.
- Reports consistent gains across simulation and real-world manipulation, including VLA, diffusion-policy, and lightweight-policy downstream settings.
- Project page and arXiv paper are live; code and Hugging Face artifacts are explicitly marked coming soon.
Research type Project PageSource project_pageAuthors Jusuk Lee, Seungjae Lee, Jonghun Shin, Hoseong Jung, Sungha Kim, Daesol Cho, H. Jin Kim, Jia-Bin Huang, Furong HuangOrganizations Seoul National University, University of Maryland, College Park, Georgia Institute of TechnologySurfacedSurfacedEvidence Strong - HumanoidMimicGen project released for whole-body humanoid loco-manipulation data generation
Full research details
Summary- Introduces a whole-body planning method for adapting contact-rich humanoid loco-manipulation skills from source demonstrations to new object states and scenes.
- Generates stable, collision-free simulated trajectories by interleaving single-arm, dual-arm, locomotion, and manipulation planning.
- Adds a nine-task simulated loco-manipulation benchmark and reports about 20 percent higher real-world success when co-training with generated data.
Research type Project PageSource project_pageAuthors Kevin Lin, Ajay Mandlekar, Caelan Reed Garrett, Nikita Chernyadev, Yu Fang, Runyu Ding, Yuqi Xie, Justin Tran, Linxi Fan, Yuke ZhuOrganizations NVIDIA, The University of Texas at AustinSurfacedSurfacedEvidence Strong - RoboGhost releases code for retargeting-free language-to-locomotion humanoid control
Full research details
Summary- Releases code, training/evaluation scripts, deployment paths, and motion-latent data for language-guided humanoid locomotion.
- Uses language-grounded motion latents to avoid explicit motion decoding and retargeting before control.
- Includes sim2sim, sim-to-real deployment, ONNX export, teacher/student policy workflows, and a motion dataset for humanoid experiments.
Research type RepoSource githubAuthors Zhe Li, Cheng Chi, Yangyang Wei, Boan Zhu, Yibo Peng, Tao Huang, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang, Chang XuOrganizations BAAI, Harbin Institute of Technology, Hong Kong University of Science and Technology, Shanghai Jiao Tong University, Peking University, University of SydneySurfacedSurfacedEvidence Strong - EXPO-FT project page and paper released for sample-efficient VLA RL finetuning
Full research details
Summary- Introduces RL finetuning for pretrained VLA policies using online interaction, Q-guided sampling, residual edit policies, and human corrections.
- Reports 30/30 success across evaluated real-robot manipulation tasks with about 19 minutes of online robot data per task.
- Project page and arXiv paper are live; code is explicitly marked coming soon, so this is tracked as a project-page/paper artifact for now.
Research type Project PageSource project_pageAuthors Perry Dong, Kuo-Han Hung, Tian Gao, Dorsa Sadigh, Chelsea FinnOrganizations Stanford UniversitySurfacedSurfacedEvidence Strong - HumanEgo releases open code for zero-shot robot learning from human egocentric videos
Full research details
Summary- Releases code for a robot-data-free, hardware-agnostic pipeline that learns manipulation policies from short human egocentric videos.
- Uses interaction-centric tokens to lift hand-object 6-DoF relationships into an embodiment-invariant representation for policy learning.
- Reports zero-shot real-robot transfer across bimanual tasks; dataset and pretrained model weights are still marked coming soon.
Research type RepoSource githubAuthors Zhi (Leo) Wang, Botao He, Kelin Yu, Seungjae Lee, Ruohan Gao, Furong Huang, Yiannis AloimonosOrganizations University of MarylandSurfacedSurfacedEvidence Strong - Flying Together shared-control framework for aerial robot teams released
Full research details
Summary- Presents a preprint on immersive shared control for aerial robot teams in unknown environments.
- Combines a VR-based operator interface with motion-primitive planning and admittance control for mixed physical and simulated drones.
- Relevant to multi-drone navigation where human guidance must shape autonomous team behavior in real time.
Research type PreprintSource arxivAuthors Lou De Bel-Air, Luca Morando, Ruitao Chen, Keru Wang, Benjamin Jarvis, Charbel Toumieh, Yang Zhou, Ken Perlin, Dario Floreano, Giuseppe LoiannoSurfacedSurfacedEvidence Strong - Closed-loop sim-to-real RL for microfiber shape control released
Full research details
Summary- Presents a preprint on closed-loop sim-to-real reinforcement learning for deformable microfiber shape control.
- Transfers an RL policy trained in a simplified simulator to a physical dual-gripper micromanipulation system using real-time visual feedback.
- Relevant to soft robotics and deformable-object manipulation where sim-to-real mismatch remains observable during control.
Research type PreprintSource arxivAuthors Alessandro Amici, Houari Bettahar, Veeti Jaakkola, Quan ZhouSurfacedSurfacedEvidence Strong - Distributed multi-coverage algorithm for robot swarms released
Full research details
Summary- Presents a preprint on distributed multi-coverage for robot swarms.
- Targets local sensing and local communication constraints where robots must maintain redundant coverage without global coordination.
- Relevant to swarm navigation and monitoring deployments that must tolerate partial robot failures.
Research type PreprintSource arxivAuthors Mariem Guitouni, Aaron T. BeckerSurfacedSurfacedEvidence Strong - rsynthlabs/sdk verifiable robot execution SDK released
Full research details
Summary- Releases an MIT-licensed SDK for signing, anchoring, and verifying robot execution payloads.
- Includes a Python SDK, Solidity contract, payload schema, CLI verifier, and an end-to-end LeRobot evaluation example.
- Relevant as open infrastructure for auditable robotics data pipelines and execution logs, though still early-stage.
Research type RepoSource githubOrganizations rsynthlabsSurfacedSurfacedEvidence Strong - HITL-D diffusion-assisted shared control paper released
Full research details
Summary- Presents a preprint on Human-In-The-Loop Diffusion for assisted shared control.
- Uses a diffusion-based policy to provide autonomous end-effector orientation updates from point-cloud scene context and Cartesian end-effector position.
- Relevant to robot manipulation and teleoperation where shared autonomy can reduce operator workload.
Research type PreprintSource arxivAuthors Riley Zilka, Sergey Khlynovskiy, Allie Wang, Martin JagersandSurfacedSurfacedEvidence Strong - PhysX-Omni simulation-ready physical 3D benchmark released
Full research details
Summary- Introduces a benchmark and dataset for simulation-ready physical 3D generation across rigid, deformable, and articulated objects.
- Releases the PhysXVerse dataset with project, code, and arXiv artifacts for evaluating physics-aware 3D asset generation.
- Useful for embodied AI and robot-learning workflows that depend on physically usable simulation assets.
Research type BenchmarkSource project_siteAuthors Yuan LiuSurfacedSurfacedEvidence Strong - Microgravity grasp-based dynamic locomotion paper released
Full research details
Summary- Presents a preprint on grasp-based dynamic locomotion for multi-limbed robots in microgravity.
- Studies gait pattern, stride length, speed, and posture choices under coupled contact and whole-body dynamics constraints.
- Relevant to space robotics systems that need locomotion by anchoring and manipulating contact points.
Research type PreprintSource arxivAuthors Chaerim Moon, Joohyung Kim, Justin K. YimSurfacedSurfacedEvidence Strong - RoboMME introduces manipulation benchmark for temporal, spatial, object, and procedural memory
Full research details
Summary- Ships a 16-task simulation benchmark covering multiple memory demands in manipulation.
- Belongs to memory-augmented robot learning benchmarks.
- Matters because it gives a cleaner way to separate different memory requirements rather than treating long-horizon failure as one monolithic issue.
Research type BenchmarkSource project_siteAuthors Yinpei Dai, Hongze Fu, Jayjun Lee, Yuejiang Liu, Haoran Zhang, Jianing Yang, Chelsea Finn, Nima Fazeli, Joyce ChaiOrganizations University of Michigan, Stanford University, Figure AISurfacedSurfacedEvidence StrongArtifact RoboMME - RoboMME introduces manipulation benchmark for temporal, spatial, object, and procedural memory
Full research details
Summary- Ships a 16-task simulation benchmark covering multiple memory demands in manipulation.
- Belongs to memory-augmented robot learning benchmarks.
- Matters because it gives a cleaner way to separate different memory requirements rather than treating long-horizon failure as one monolithic issue.
Research type BenchmarkSource project_siteAuthors Yinpei Dai, Hongze Fu, Jayjun Lee, Yuejiang Liu, Haoran Zhang, Jianing Yang, Chelsea Finn, Nima Fazeli, Joyce ChaiOrganizations University of Michigan, Stanford University, Figure AISurfacedSurfacedEvidence StrongArtifact RoboMME - AllenAI releases bimanual pegboard VLA evaluation dataset
Full research details
Summary- Releases a LeRobot-compatible bimanual pegboard manipulation dataset on Hugging Face.
- Includes 52 episodes, roughly 30.5k frames, 14-DoF bimanual arm state/action traces, and video/time-series data for in-distribution evaluation.
- Useful for reproducible VLA benchmarking on bimanual tabletop manipulation with standardized dataset packaging.
Research type DatasetSource huggingfaceOrganizations Allen Institute for AISurfacedSurfacedEvidence Strong - RoboGen generative simulation pipeline for robot learning released
Full research details
Summary- Official ICML 2024 repository for a self-guided generative robot-learning pipeline.
- Reimplements task proposal, scene generation, and skill learning in PyBullet for rigid manipulation and locomotion tasks.
- Includes generated-task workflows, training scripts, logs, visualizations, and Apache-2.0 code for automated robot learning research.
Research type RepoSource githubAuthors Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, Chuang GanSurfacedSurfacedEvidence Strong - VLA-REPLICA real-world VLA benchmark released
Full research details
Summary- Introduces a benchmark for low-cost, reproducible real-world evaluation of Vision-Language-Action models.
- Uses an SO-101 arm, depth camera, wrist webcam, and lightbox setup with standardized manipulation tasks and evaluation protocols.
- Useful for cross-lab VLA testing where reproducible hardware and task conditions matter.
Research type BenchmarkSource project_siteAuthors Alex S. Huang, Jiahui Zhang, Shiqing Tang, Yu XiangSurfacedSurfacedEvidence Strong - RoboRacer trajectory planning and control benchmark released
Full research details
Summary- Introduces a benchmark for high-speed trajectory planning and control on the RoboRacer platform.
- Includes raceline generation, online velocity replanning, path-tracking controllers, and inverse-dynamics learning for aggressive driving.
- Relevant to autonomous racing and navigation research where policies operate near physical limits.
Research type BenchmarkSource arxivAuthors Mattia Piccinini, Patrick Zambiasi, Aniello Mungiello, Mattia Piazza, Felix Jahncke, Johannes BetzSurfacedSurfacedEvidence Strong - Multimodal Diffusion Forcing for forceful robot manipulation released
Full research details
Summary- Introduces a diffusion-forcing model over multimodal robot trajectories for contact-rich manipulation.
- Uses random partial masking to learn temporal and cross-modal dependencies among observations, actions, rewards, and force signals.
- Demonstrates policy, dynamics, world/action modeling, and anomaly-detection behavior on simulated and real forceful manipulation tasks.
Research type PreprintSource arxivAuthors Zixuan Huang, Huaidian Hou, Dmitry BerensonOrganizations University of MichiganSurfacedSurfacedEvidence Strong - TienKung-Lab humanoid locomotion training stack released
Full research details
Summary- Releases a repository for TienKung humanoid locomotion training in Isaac Lab.
- Combines RL, AMP-style rewards, periodic gait rewards, ray-casting sensors, and Sim2Sim transfer to MuJoCo.
- Relevant to humanoid locomotion stacks that need reproducible simulation training and real-robot validation.
Research type RepoSource githubOrganizations Open-X-HumanoidSurfacedSurfacedEvidence Strong - MIKASA-Robo-VLA memory benchmark for robotic manipulation released
Full research details
Summary- Extends MIKASA-Robo into a VLA benchmark for memory-intensive tabletop manipulation.
- Provides 90 registered environments across 10 memory types, language instructions, horizon splits, and standardized evaluation protocols.
- Releases trajectories in RLDS and LeRobotDataset formats for reproducible VLA training and evaluation.
Research type BenchmarkSource project_siteAuthors Egor Cherepanov, Nikita Kachaev, Alexey Kovalev, Aleksandr PanovSurfacedSurfacedEvidence Strong - SimToolReal zero-shot dexterous tool manipulation released
Full research details
Summary- Introduces an object-centric policy for zero-shot dexterous manipulation of unseen tools and tasks.
- Trains over procedurally generated tool-like primitives in simulation, then evaluates on real tools across multiple categories and task trajectories.
- Releases a project page with paper, code, interactive simulation demo, and DexToolBench assets for dexterous tool-use evaluation.
Research type Project PageSource project_siteAuthors Kushal Kedia, Tyler Ga Wei Lum, Jeannette Bohg, C. Karen LiuOrganizations Stanford UniversitySurfacedSurfacedEvidence Strong - LeRobot Humanoid open low-cost biped platform released
Full research details
Summary- Project page for an open, low-cost humanoid robot platform aimed at robot-learning experiments.
- Releases hardware build assets, assembly documentation, runtime tools, simulator identification, and training environments around a 3D-printed biped.
- Useful for tracking accessible humanoid platforms that close the loop from hardware design to real-world control.
Research type Project PageSource githubOrganizations Hugging FaceSurfacedSurfacedEvidence Strong - AllenAI releases MolmoAct2 action reasoning model and bimanual dataset
Full research details
Summary- Introduces a fully open action reasoning model for real-world robot deployment.
- Releases model weights, training code, OpenFAST action tokenizer, and new robotics datasets including a 720-hour bimanual manipulation corpus.
- Targets practical VLA deployment across low- and medium-cost robot platforms, with evaluations spanning simulation, real-world control, and embodied reasoning benchmarks.
Research type Model ReleaseSource huggingfaceOrganizations Allen Institute for AISurfacedSurfacedEvidence StrongArtifact MolmoAct2 - ViserDex visual sim-to-real dexterous reorientation released
Full research details
Summary- Introduces a visual sim-to-real framework for monocular RGB dexterous in-hand reorientation.
- Uses 3D Gaussian Splatting domain randomization for object pose estimation, plus curriculum RL and teacher-student distillation for control.
- Validates robust real-world transfer on a physical multi-fingered hand under varied lighting, with training feasible on consumer-grade hardware.
Research type Project PageSource project_siteAuthors Arjun Bhardwaj, Maximum Wilder-Smith, Mayank Mittal, Vaishakh Patil, Marco HutterOrganizations ETH Zurich, NVIDIASurfacedSurfacedEvidence Strong - DexJoCo releases benchmark and toolkit for dexterous manipulation in MuJoCo
Full research details
Summary- Benchmark and toolkit for task-oriented dexterous manipulation in MuJoCo.
- Includes dexterous hand tasks covering tool use, bimanual coordination, long-horizon execution, and reasoning.
- Provides code, demonstrations, and evaluation tooling for imitation learning and VLA benchmarking.
Research type BenchmarkSource webAuthors Hanwen Wang, Weizhi Zhao, Xiangyu Wang, Siyuan Huang, He Lin, Boyuan Zheng, Rongtao Xu, Gang Wang, Yao Mu, He Wang, Lue Fan, Hongsheng Li, Zhaoxiang Zhang, Tieniu TanOrganizations CASIA, Shanghai Jiao Tong University, MBZUAI, Peking University, Galbot, CUHKSurfacedSurfacedEvidence Strong - DiT4DiT releases real-time whole-body control code for humanoid robots
Full research details
Summary- Open-source implementation for real-time whole-body control of humanoid robots.
- Uses a diffusion-transformer policy for coordinated humanoid locomotion and manipulation.
- Supports humanoid control and embodied AI research pipelines.
Research type RepoSource githubOrganizations Mondo RoboticsSurfacedSurfacedEvidence Strong - Hand-in-the-Loop introduces interventional correction for dexterous VLA policies
Full research details
Summary- Project release for Hand-in-the-Loop, a method for corrective human intervention in dexterous VLA policies.
- Targets bimanual coordination, tool use, and long-horizon manipulation with high-DoF hands.
- Frames intervention as a way to improve data collection, correction, and policy refinement for dexterous manipulation.
Research type Project PageSource webAuthors Zhuohang Li, Liqun Huang, Wei Xu, Zhengming Zhu, Nie Lin, Xiao Ma, Xinjun Sheng, Ruoshi WenOrganizations Shanghai Jiao Tong University, ByteDance Seed, University of TokyoSurfacedSurfacedEvidence Strong - Lichtblick releases browser-based robotics diagnostics visualizer
Full research details
Summary- Browser-native robotics diagnostics and visualization tool for live robot telemetry.
- Supports debugging, monitoring, and robotics development workflows with real-time data streams.
- Useful as developer infrastructure for robotics software stacks.
Research type RepoSource githubSurfacedSurfacedEvidence StrongArtifact lichtblick - MathWorks publishes large open-source robotics simulation repository
Full research details
Summary- Large open-source robotics simulation and control repository covering robot arms, drones, AGVs, ROS workflows, and navigation environments.
- Includes simulation examples, deployment workflows, and robotics tooling for education and prototyping.
- Functions as broad robotics infrastructure and simulation reference material.
Research type RepoSource githubOrganizations MathWorksSurfacedSurfacedEvidence Strong - GMR releases CPU-based humanoid motion retargeting framework
Full research details
Summary- Open-source framework for retargeting human motion capture to humanoid robot poses in real time on CPU hardware.
- Supports teleoperation and imitation learning workflows for humanoid control.
- Useful as lightweight infrastructure for whole-body humanoid motion pipelines.
Research type RepoSource githubSurfacedSurfacedEvidence StrongArtifact GMR – Gaussian Motion Retargeting - ACWM-Phys releases physics-rich benchmark for action-conditioned world models
Full research details
Summary- Benchmark and dataset release for evaluating action-conditioned world models across multiple physics regimes.
- Covers rigid contact, deformable interaction, particle dynamics, and kinematic systems with controlled generalization splits.
- Includes datasets, checkpoints, evaluation tooling, and diffusion-transformer baselines for embodied world-model research.
Research type BenchmarkSource arxivAuthors Haotian Xue, Yongxin Chen, et al.Organizations Georgia Tech, MetaSurfacedSurfacedEvidence Strong - Spes Robotics releases phone-based robot arm teleoperation framework
Full research details
Summary- Open-source framework enabling smartphone-based teleoperation for robot arms.
- Provides accessible real-time control workflows for manipulation systems.
- Useful for low-cost teleoperation, data collection, and imitation learning experiments.
Research type RepoSource githubOrganizations Spes RoboticsSurfacedSurfacedEvidence Strong - Genesis-World releases open-source physics engine for robotics simulation
Full research details
Summary- Open-source physics engine for robotics simulation and embodied AI workflows.
- Targets scalable training and evaluation pipelines for robot learning.
- Useful as simulation infrastructure, but source engagement is low and artifact quality should be checked directly.
Research type RepoSource github.comOrganizations Genesis Embodied AISurfacedSurfacedEvidence MediumArtifact genesis-world - KLEIYN-EFGCL introduces energy-based learning for quadruped locomotion
Full research details
Summary- Preprint and project release for energy-based flow-guided contrastive learning in quadruped locomotion.
- Focuses on improving sample efficiency and policy learning for legged robots.
- Provides project materials for reproducing the quadruped locomotion method.
Research type PreprintSource arxiv.orgAuthors Keita Yoneda, et al.Organizations University of Tokyo, JSK Robotics LaboratorySurfacedSurfacedEvidence Strong - RoboVerse releases cross-embodiment robot learning framework
Full research details
Summary- Open-source framework for cross-embodiment robot learning across simulator environments.
- Provides training and evaluation infrastructure for comparing policies across robot forms.
- Relevant to embodied AI workflows where transfer across embodiments and simulators is the core problem.
Research type RepoSource github.comOrganizations RoboVerseOrgSurfacedSurfacedEvidence MediumArtifact RoboVerse - LaST-R1 releases latent reasoning framework for VLA robot policies
Full research details
Summary- Project release for LaST-R1, a VLA framework using latent physical reasoning before action generation.
- Combines adaptive reasoning steps with reinforcement-style post-training for manipulation policies.
- Targets improved generalization and policy refinement across single-arm and dual-arm robot tasks.
Research type Project PageSource arxiv.orgAuthors Hao Chen, Jiaming Liu, Zhonghao Yan, Nuowei Han, Renrui Zhang, Chenyang Gu, Jialin Gao, Ziyu Guo, Siyuan Qian, Yinxi Wang, Peng Jia, Chi-Wing Fu, Shanghang Zhang, Pheng-Ann HengOrganizations The Chinese University of Hong Kong, Peking University, Simplexity RoboticsSurfacedSurfacedEvidence Strong - Axis Robotics releases data-cleaning pipeline for robot demonstration trajectories
Full research details
Summary- Open-source pipeline for converting noisy robot demonstration trajectories into model-ready training data.
- Targets browser teleoperation and crowdsourced demonstration workflows with variable timing and motion artifacts.
- Supports imitation learning and physical AI data generation pipelines.
Research type RepoSource github.comOrganizations Axis RoboticsSurfacedSurfacedEvidence StrongArtifact AxisDataCleaning - DexSim2Real proposes foundation-model-guided sim-to-real transfer for dexterous manipulation
Full research details
Summary- Preprint on foundation-model-guided sim-to-real transfer for dexterous manipulation.
- Focuses on improving generalization from simulation to real robotic hardware.
- Paper-only artifact with no code or dataset surfaced in the provided record.
Research type PreprintSource arxiv.orgAuthors Zijian Zeng, Fei Ding, Huiming Yang, Xianwei Li, Yuhao LiaoSurfacedSurfacedEvidence Medium - TeleAI releases PRTS policy framework for humanoid decision-making
Full research details
Summary- Open-source policy framework for humanoid decision-making with test-time scaling.
- Bridges imitation-style policy learning with planning-oriented decision refinement.
- Robotics relevance is explicit, but the source signal is very weak and should be reviewed carefully.
Research type RepoSource github.comOrganizations TeleAISurfacedSurfacedEvidence MediumArtifact PRTS - XLeRobot releases low-cost open-source dual-arm mobile robot platform
Full research details
Summary- Open-source dual-arm mobile robot platform with hardware and software materials.
- Targets accessible mobile manipulation experimentation at low cost.
- Useful as a builder platform rather than a core research method or benchmark.
Research type RepoSource github.comAuthors Vector-WangelSurfacedSurfacedEvidence MediumArtifact XLeRobot - Google DeepMind maintains MuJoCo Menagerie robot model library
Full research details
Summary- Curated library of robot models for MuJoCo simulation.
- Provides ready-to-use MJCF models across multiple robot embodiments.
- Useful as simulation infrastructure for reinforcement learning, control, and sim-to-real research.
Research type RepoSource github.comOrganizations Google DeepMindSurfacedSurfacedEvidence MediumArtifact mujoco_menagerie - RLWRLD releases RLDX-1 dexterity-first robotics foundation model
Full research details
Summary- Releases RLDX-1, a robotics foundation model focused on dexterous manipulation.
- Uses multimodal action modeling across vision, language, tactile, force, and torque signals.
- Includes model checkpoints and dexterity evaluation material for manipulation-focused embodied AI research.
Research type Model ReleaseSource huggingface.coAuthors RLWRLD research teamOrganizations RLWRLDSurfacedSurfacedEvidence StrongArtifact RLDX-1 - GS-Playground releases photorealistic 3DGS simulator for robot learning
Full research details
Summary- Open-source 3D Gaussian Splatting simulation framework for vision-informed robot learning.
- Combines photorealistic rendering, physics simulation, and real-to-sim tooling for embodied policy training.
- Supports high-throughput simulation workflows for navigation and manipulation research.
Research type Project PageSource arxiv.orgAuthors Yufei Jia, et al.Organizations Tsinghua UniversitySurfacedSurfacedEvidence Strong - NVIDIA releases SONIC whole-body control stack through GR00T-WholeBodyControl
Full research details
Summary- Open-source whole-body control stack for humanoid locomotion and manipulation research.
- Includes teleoperation, data collection, planning, inference, and policy training components.
- Targets community humanoid platforms and integration with GR00T-style VLA post-training workflows.
Research type RepoSource github.comAuthors Zhengyi Luo, et al.Organizations NVIDIASurfacedSurfacedEvidence StrongArtifact GR00T-WholeBodyControl - SCFields introduces semantic-contact representations for tool manipulation
Full research details
Summary- Introduces semantic-contact fields combining object semantics with contact-aware geometry for manipulation.
- Targets category-level generalization for tool use across unseen objects.
- Provides code and evaluation for learning contact-aware representations in manipulation tasks.
Research type Project PageSource otherAuthors Kevin Ma, et al.Organizations National University of SingaporeSurfacedSurfacedEvidence Strong - Lucid-XR introduces XR-based data engine for robotic manipulation
Full research details
Summary- XR-based data engine for generating manipulation training data.
- Focuses on scalable data creation for embodied AI systems.
- Provides project page and preprint for data-driven robotics workflows.
Research type Project PageSource otherAuthors Yajvan Ravan, et al.SurfacedSurfacedEvidence Medium - ObotX releases motion planning package for mobile manipulators
Full research details
Summary- Motion planning and ROS integration package for mobile manipulation.
- Supports coordinated base and arm trajectory generation.
- Targets open-source mobile manipulation platforms.
Research type RepoSource githubOrganizations ObotXSurfacedSurfacedEvidence Medium - Asimov releases motion control board for open-source humanoid platform
Full research details
Summary- Open-source motion control board completing the Asimov v1 humanoid hardware stack.
- Provides motor and sensor interfacing, onboard IMU, and integration with compute modules.
- Enables full-stack reproducibility for humanoid hardware and control experiments.
Research type RepoSource githubOrganizations AsimovSurfacedSurfacedEvidence StrongArtifact Asimov v1 Motion Control Board - OmniStream releases real-time perception-to-manipulation framework
Full research details
Summary- Open-source framework combining video perception and geometric reconstruction for manipulation.
- Connects scene understanding directly to action execution pipelines.
- Targets real-time robotic manipulation workflows.
Research type RepoSource githubSurfacedSurfacedEvidence MediumArtifact OmniStream - KinDER introduces a benchmark for physical reasoning in robot learning
Full research details
Summary- Benchmark for evaluating physical reasoning in robot learning and planning.
- Focuses on tasks requiring physical commonsense understanding.
- Provides evaluation setup for reasoning-driven robotics systems.
Research type PreprintSource arxivAuthors Y. Huang, et al.Organizations Princeton University, Carnegie Mellon University, Georgia TechSurfacedSurfacedEvidence Strong - NVIDIA releases ProtoMotions humanoid simulation framework
Full research details
Summary- GPU-accelerated simulation framework for humanoid robot training and evaluation.
- Designed for scalable policy learning in simulated environments.
- Supports humanoid locomotion and control research workflows.
Research type RepoSource githubOrganizations NVIDIASurfacedSurfacedEvidence Strong - A2A introduces single-step flow matching for real-time robot action generation
Full research details
Summary- Project release for Action-to-Action Flow Matching, a policy approach for fast robot action generation.
- Uses the previous action as the starting point for the next action prediction, reducing inference steps for real-time control.
- Targets manipulation policies where low-latency action generation and robustness to visual variation matter.
Research type Project PageSource otherAuthors Jindou Jia, Gen Li, Xiangyu Chen, Tuo An, Yuxuan Hu, Jingliang Li, Xinying Guo, Jianfei YangSurfacedSurfacedEvidence StrongArtifact Action-to-Action Flow Matching - Squint releases fast visual reinforcement learning code for sim-to-real robotics
Full research details
Summary- Open-source implementation of a visual reinforcement learning method for sim-to-real robot manipulation.
- Uses parallelized training, image preprocessing, and domain randomization to improve wall-clock learning speed.
- Includes training and deployment tooling for low-cost robot arm experiments.
Research type RepoSource githubAuthors Abdulaziz Almuzairee, Henrik I. ChristensenOrganizations UC San DiegoSurfacedSurfacedEvidence Strong - OmniRobotHome introduces a multi-camera platform for human-robot interaction
Full research details
Summary- Research platform for studying real-time multi-human and multi-robot interaction in home-like environments.
- Uses a room-scale multi-camera setup with robot arms to capture aligned human, object, and robot activity.
- Supports perception and interaction research around shared-space human-robot collaboration.
Research type Project PageSource otherAuthors Junyoung Lee, Sookwan Han, Jeonghwan Kim, Inhee Lee, Mingi Choi, Jisoo Kim, Wonjung Woo, Hanbyul JooOrganizations Seoul National UniversitySurfacedSurfacedEvidence Strong - LaST-R1 introduces physical latent reasoning for VLA robot policies
Full research details
Summary- Preprint proposing adaptive physical latent reasoning for reinforcement in vision-language-action robot policies.
- Focuses on improving robotic manipulation by reasoning over latent physical action structure.
- No code or dataset release is included in the provided artifact.
Research type PreprintSource arxivAuthors Hao Chen, Jiaming Liu, Zhonghao Yan, Nuowei Han, Renrui Zhang, Chenyang Gu, Jialin Gao, Ziyu Guo, Siyuan Qian, Yinxi Wang, Peng Jia, Chi-Wing Fu, Shanghang Zhang, Pheng-Ann HengSurfacedSurfacedEvidence Medium - FlexiTac releases an open-source tactile sensing platform for robotic end-effectors
Full research details
Summary- Open-source tactile sensing platform for robotic grippers, hands, and end-effectors.
- Uses flexible tactile sensor pads and compact readout hardware for dense contact sensing.
- Supports contact-rich manipulation research and scalable tactile data collection.
Research type Project PageSource otherAuthors Binghao Huang, Yunzhu LiSurfacedSurfacedEvidence Strong - FreeOcc proposes training-free open-vocabulary occupancy prediction for embodied agents
Full research details
Summary- Preprint on training-free open-vocabulary 3D occupancy prediction for embodied perception.
- Targets scene understanding from visual sequences without task-specific 3D annotation or training.
- Robotics relevance is strongest for navigation, mapping, and embodied spatial perception.
Research type PreprintSource arxivAuthors Zeyu Jiang, Changqing Zhou, Xingxing Zuo, Changhao ChenSurfacedSurfacedEvidence Medium - DAIMON Robotics releases Daimon-Infinity omni-modal robotics dataset
Full research details
Summary- Dataset release for omni-modal robot manipulation data with tactile, visual, force, and interaction signals.
- Focuses on fine manipulation scenarios where contact, deformation, slip, and material properties matter.
- Supports training and evaluation of manipulation models that need richer sensing than vision alone.
Research type DatasetSource otherOrganizations Daimon RoboticsSurfacedSurfacedEvidence StrongArtifact Daimon-Infinity - LARYBench introduces a benchmark for latent action representations in robotic video
Full research details
Summary- Large-scale benchmark and dataset for evaluating latent action representations from video in robotics.
- Includes diverse video, trajectory, and multi-robot data for testing action understanding and motion prediction.
- Provides standardized evaluation tasks and tooling for studying representation quality in manipulation and control pipelines.
Research type BenchmarkSource otherOrganizations MeituanSurfacedSurfacedEvidence Strong - NVIDIA releases GR00T-VisualSim2Real pipeline for humanoid sim-to-real transfer
Full research details
Summary- Open-source pipeline for visual sim-to-real transfer of humanoid locomotion and manipulation skills.
- Includes training code, simulation assets, and workflows for learning behaviors such as walking and object interaction.
- Targets scalable deployment of humanoid policies without reliance on real-world teleoperation data.
Research type RepoSource githubOrganizations NVIDIASurfacedSurfacedEvidence Strong - AIST releases RoboManipBaselines framework for imitation learning in robotic manipulation
Full research details
Summary- Open-source framework integrating imitation learning methods and benchmark environments for robotic manipulation.
- Supports policy training, teleoperation data collection, and preprocessing for vision and point cloud inputs.
- Provides baseline implementations and evaluation pipelines across common manipulation tasks.
Research type RepoSource githubAuthors Masaki Murooka, Tomohiro Motoda, Ryoichi Nakajo, Hanbit Oh, Koshi Makihara, Keisuke Shirai, Tetsuya Ogata, Yukiyasu DomaeOrganizations AISTSurfacedSurfacedEvidence Strong - NVIDIA and Stanford release PointWorld 3D world model training pipeline
Full research details
Summary- Open-source training and evaluation pipeline for PointWorld, a 3D world model for robotic manipulation.
- Predicts scene-level 3D point flows from RGB-D observations and robot actions.
- Supports training and evaluation across datasets such as DROID and BEHAVIOR for real-world manipulation scenarios.
Research type RepoSource githubAuthors Wenlong Huang, Yu-Wei Chao, Arsalan Mousavian, Ming-Yu Liu, Dieter Fox, Kaichun Mo, Li Fei-FeiOrganizations NVIDIA, Stanford UniversitySurfacedSurfacedEvidence Strong - TetherIA releases Aero Hand Open tendon-driven robotic hand platform
Full research details
Summary- Open-source tendon-driven robotic hand with 7 degrees of freedom for dexterous manipulation research.
- Includes full hardware stack with CAD files, electronics, firmware, and ROS2-compatible software tooling.
- Designed as a low-cost, reproducible platform for experimentation with manipulation and teleoperation tasks.
Research type RepoSource githubOrganizations TetherIASurfacedSurfacedEvidence Strong - Axis Robotics releases sample dataset for long-horizon manipulation tasks
Full research details
Summary- Sample dataset of long-horizon manipulation trajectories featuring a Franka arm performing a kettle pouring task.
- Includes raw and processed trajectory variants with preprocessing and structure details.
- Serves as a reference dataset for imitation learning and physical AI data pipeline development.
Research type DatasetSource otherOrganizations Axis RoboticsSurfacedSurfacedEvidence Strong - NVIDIA releases MotionBricks generative motion framework for robotics
Full research details
Summary- Generative motion framework for real-time control using a modular latent representation of skills.
- Supports composition of locomotion, object interaction, and scene interaction behaviors without retraining.
- Demonstrated on humanoid systems for whole-body locomotion and manipulation tasks.
Research type Project PageSource otherAuthors Tingwu Wang, et al.Organizations NVIDIASurfacedSurfacedEvidence Strong - Xiaomi Robotics open-sources Xiaomi-Robotics-0 VLA post-training pipeline
Full research details
Summary- Open-source release of Xiaomi-Robotics-0, a vision-language-action model with post-training code and model weights.
- Includes data processing, training scripts, attention masking, action prefixing, and policy fine-tuning components.
- Targets cross-embodiment manipulation research across simulation benchmarks and real bimanual tasks.
Research type Model ReleaseSource githubAuthors Cai Rui, Guo Jun, He Xinze, Jin Piaopiao, Li Jie, Lin Bingxuan, Liu Futeng, Liu Wei, Ma Fei, Ma Kun, Qiu Feng, Qu Heng, Su Yifei, Sun Qiao, Wang Dong, Wang Donghao, Wang Yunhong, Wu Rujie, Xiang Diyun, Yang Yu, Ye Hangjun, Zhang Yuan, Zhou QuanyunOrganizations Xiaomi RoboticsSurfacedSurfacedEvidence Strong - Meta releases SPIDER framework for physics-informed dexterous retargeting
Full research details
Summary- Open-source release of SPIDER, a physics-informed framework for retargeting human motion to robot embodiments.
- Supports dexterous hands, humanoids, and multiple simulator backends for manipulation and policy training workflows.
- Provides tooling for contact-aware retargeting, simulation data generation, and downstream robot learning pipelines.
Research type RepoSource githubAuthors Chaoyi Pan, Changhao Wang, Haozhi Qi, Zixi Liu, Homanga Bharadhwaj, Akash Sharma, Tingfan Wu, Guanya Shi, Jitendra Malik, Francois HoganOrganizations MetaSurfacedSurfacedEvidence Strong