Physical Intelligence releases FAST, a robot action tokenizer enabling 5x faster VLA training

Physical Intelligence published FAST, a robot action tokenizer that enables training of autoregressive vision-language-action models on complex, dexterous manipulation tasks that prior tokenization approaches fail on. Using FAST, Physical Intelligence trained π0-FAST, a generalist robot policy that matches the performance of its diffusion-based π0 model across dexterous and long-horizon manipulation tasks while reducing training time by up to 5x.

The Company

Physical Intelligence was founded in early 2024 in San Francisco. CEO Karol Hausman previously worked as a Staff Research Scientist at Google DeepMind focusing on manipulation and locomotion algorithms. Chief Scientist Sergey Levine is an associate professor at UC Berkeley whose lab pioneered deep reinforcement learning for robot manipulation and sim-to-real transfer. Research Lead Chelsea Finn is an associate professor at Stanford known for her work on model-agnostic meta-learning, which underpins how robots can adapt to new tasks with minimal data. Co-founders Brian Ichter and Lachy Groom, previously at Google and Stripe respectively, round out a team that is unusually academic in composition for a venture-backed startup.

Physical Intelligence raised $70 million in seed funding in March 2024, $400 million in Series A in November 2024 at a $2 billion valuation with Jeff Bezos, OpenAI, Thrive Capital, and Lux Capital participating, and $600 million in Series B in November 2025 led by CapitalG, Alphabet's growth fund, at a $5.6 billion valuation; total funding now exceeds $1.1 billion.

The company's explicit position is that it does not manufacture robots. It builds the software and model layer and trains across hardware platforms built by others. Physical Intelligence's training infrastructure spans seven different robot platforms. FAST being described as a universal tokenizer is not incidental to that strategy; a tokenizer that works across single-arm, dual-arm, and mobile robots is the infrastructure layer a hardware-agnostic company needs.

The Problem

Most foundation models use Transformer architectures that operate on discrete tokens. For language, this works because byte-pair encoding compresses text efficiently. For robot control, the equivalent step is tokenizing continuous action signals, and the standard approach of simple per-dimension, per-timestep binning performs poorly on dexterous tasks from high-frequency robot data. The underlying issue is that action tokens produced by binning are highly correlated across time, which weakens the next-token prediction objective that autoregressive VLAs rely on. Training on dexterous data with binned tokens either fails entirely or requires diffusion-based architectures that are more computationally expensive to train.

What FAST Does

FAST applies time-series compression to robot action trajectories, removing temporal correlation before tokenization and allowing the autoregressive prediction objective to function effectively on high-frequency control signals. The FAST+ variant is a universal tokenizer trained on one million real robot action trajectories and works across single-arm, dual-arm, and mobile robots with different action dimensions and control frequencies, functioning as an off-the-shelf tokenizer without dataset-specific training. Physical Intelligence also released code allowing users to train custom FAST tokenizers on their own data in minutes.

When integrated with π0, FAST-based autoregressive VLAs scale to training on 10,000 hours of robot data. The 5x training speedup is meaningful beyond compute cost; faster training cycles make it more practical to iterate on dexterous manipulation policies at research and commercial scale.

Open Source

On February 4, 2025, Physical Intelligence released the weights and code for both π0 and π0-FAST. The models and FAST tokenizer are available on Hugging Face. Publishing both a foundational method and a trained model is consistent with the academic culture Levine and Finn brought from Berkeley and Stanford; the open release builds the ecosystem of researchers and developers that a hardware-agnostic software company benefits from.

Maturity

FAST is a research contribution with open-source weights; it is not a commercial product. Physical Intelligence evaluates its models across manipulation tasks on physical hardware but has not disclosed commercial deployments or revenue. The company's stated direction is general-purpose physical intelligence models; π0 and π0-FAST are the first public milestones on that roadmap.

Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to tips@korthos.xyz or reach out on X at @agkorthos.