Agility Research

Research

  • Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning - 2025-11-06 - Research Publication - The study presents Isaac Lab, the natural successor to Isaac Gym, which extends the paradigm of GPU-native robotics simulation into the era of large-scale multi-modal learning. Digit is used in the study's data-collection… - source: arxiv.org
  • No More Marching: Learning Humanoid Locomotion for Short-Range SE(2) Targets - 2025-08-16 - Research Publication - Humanoids operating in real-world workspaces must frequently execute task-driven, short-range movements to SE(2) target poses. To be practical, these transitions must be fast, robust, and energy efficient. Digit is the… - source: arxiv.org
  • Dynamic subgoal pursuit paper validates humanoid navigation using Digit simulation - 2025-06-02 - Research Publication - The paper proposes a hierarchical navigation framework combining a reinforcement-learning subgoal planner with MPC-based bipedal walking. The authors validate the method in simulation using the Agility Robotics Digit… - source: arxiv.org
  • PPF paper improves Digit locomotion with model-assumption-based regularization - 2025-04-14 - Research Publication - PPF combines controller imitation, reinforcement-learning fine-tuning, and model-assumption-based regularization for humanoid locomotion. The paper reports hardware experiments on full-size Digit, including 1.5 m/s walking… - source: arxiv.org
  • Challenging-terrain locomotion paper trains Digit for real-world outdoor walking - 2024-10-04 - Research Publication - The paper presents a transformer-based learning approach for blind humanoid locomotion over rough, deformable, and sloped terrain. The authors evaluate the controller on a real Digit humanoid, including hiking trails in… - source: arxiv.org
  • Digit system-identification paper improves constrained-robot model tracking - 2024-08-16 - Research Publication - The paper introduces an iterative least-squares system-identification approach for constrained robots such as legged humanoids. The approach is validated in simulation and on real Agility Robotics Digit hardware… - source: arxiv.org
  • Multi-biped payload-transport paper demonstrates coordinated Cassie teams - 2024-06-25 - Research Publication - The paper trains a decentralized controller for payload transport using multiple rigidly attached biped robots. The authors demonstrate real-world transport systems composed of two and three Cassie robots. It adds a… - source: arxiv.org
  • Digit social-navigation paper combines MPC with zonotope neural networks - 2024-03-25 - Research Publication - The paper proposes cascaded zonotope-based neural networks for pedestrian prediction and bipedal social path planning. The learned social planner is integrated with MPC footstep planning for Agility Robotics Digit in… - source: arxiv.org
  • Next-token prediction paper demonstrates real-world humanoid locomotion on Digit - 2024-02-29 - Research Publication - The paper casts humanoid locomotion as next-token prediction over sensorimotor trajectories using a causal transformer. The authors use Agility Robotics Digit as the physical humanoid platform and report zero-shot… - source: arxiv.org
  • IJRR study demonstrates broad dynamic locomotion skills on Cassie - 2024-01-30 - Research Publication - The IJRR study presents a reinforcement-learning controller for standing, walking, running, jumping, and other dynamic bipedal skills. The resulting policies deploy on Cassie in extensive real-world experiments, including… - source: arxiv.org
  • Safe whole-body task-space control paper validates hardware experiments on Digit - 2023-10-12 - Research Publication - The paper proposes inverse-dynamics whole-body control with exponential control barrier functions for safety-critical humanoid behaviors. The authors apply the method to the 3D biped robot Digit in both simulation and… - source: arxiv.org
  • Cassie constrained-foothold paper demonstrates MPC footstep control - 2023-09-14 - Research Publication - The paper introduces a mixed-integer model-predictive footstep controller for bipedal walking over constrained footholds. Hardware experiments on Cassie use real-time elevation mapping and convex terrain decomposition to… - source: arxiv.org
  • OP3 paper demonstrates global-position tracking for multi-domain bipedal walking - 2023-04-27 - Research Publication - The paper introduces a stabilizing global-position tracking controller for hybrid multi-domain bipedal walking. Simulations on ROBOTIS OP3 cover multi-domain walking with different actuation regimes, walking speeds, and… - source: arxiv.org
  • RSS jumping-control paper transfers versatile dynamic jumps to Cassie hardware - 2023-02-19 - Research Publication - The RSS paper trains a reinforcement-learning controller for robust bipedal jumping across different distances and directions. After multi-stage training, the policy transfers directly to real Cassie hardware for standing… - source: arxiv.org
  • Cassie ALIP-MPC paper demonstrates terrain-adaptive bipedal locomotion - 2021-09-30 - Research Publication - The paper combines an Angular Momentum Linear Inverted Pendulum model with MPC foot placement and virtual constraints for terrain-adaptive walking. Experiments validate Cassie locomotion across surfaces with varied… - source: arxiv.org
  • Digit paper learns robust bipedal walking with reinforcement learning - 2021-03-29 - Research Publication - The paper presents robust feedback motion policy design for Agility Robotics Digit using reinforcement learning. The controller transfers from simulation to physical Digit hardware and realizes sustained walking under… - source: arxiv.org
  • Parameterized locomotion RL paper transfers robust walking policies to Cassie - 2021-03-26 - Research Publication - The paper trains robust locomotion policies in simulation using domain randomization and transfers them to the physical Cassie robot. The learned policies support dynamic behaviors such as velocity tracking, walking-height… - source: arxiv.org
  • Cassie sim-to-real paper learns all common bipedal gaits - 2020-11-02 - Research Publication - The paper proposes periodic reward composition for standing, walking, hopping, running, and skipping without requiring reference motions. The authors demonstrate sim-to-real transfer of learned gaits to Cassie and a… - source: arxiv.org
  • Cassie Hovershoes paper demonstrates multimodal bipedal riding control - 2019-07-26 - Research Publication - The paper develops feedback control for Cassie to autonomously ride a pair of Hovershoes while balancing and regulating forward and rotational velocity. Experiments include flat terrain, slopes, stairs, rough outdoor… - source: arxiv.org
  • Deep-RL policy-transfer paper reports hardware results on ATRIAS - 2018-09-28 - Research Publication - The paper studies whether deep reinforcement-learning policies can transfer from simulation to hardware when embedded inside a structured walking controller. Experiments on ATRIAS examine action spaces, cost functions, and… - source: arxiv.org
  • Cassie feedback-control paper demonstrates walking, standing, and Segway riding - 2018-09-19 - Research Publication - The paper implements virtual-constraint and gait-library feedback control on Agility Robotics Cassie. The controller enabled standing and walking across sidewalks, grass, snow, sand, and burning brush, plus a Segway-riding… - source: arxiv.org
  • Bayesian-optimization study tunes walking controllers on ATRIAS hardware - 2017-09-18 - Research Publication - The paper applies Bayesian optimization with a domain-informed feature transform to bipedal locomotion controller tuning. The method is evaluated on the ATRIAS biped in simulation and on hardware, including two walking… - source: arxiv.org

Articles

View Agility overview