Physical Intelligence Pi-0-FAST

Par Physical Intelligence (PI)

Dibināts 2024 · San Francisco, California, USA

Physical Intelligence (PI) was founded in 2024 in San Francisco by Sergey Levine, Chelsea Finn, Karol Hausman and other co-founders, with a mission to build foundation models for general-purpose robots. π-0-FAST, released in early 2025, is the autoregressive variant of the π-0 VLA introduced together with the FAST action tokenizer. FAST (Frequency-space Action Sequence Tokenization) uses Discrete Cosine Transform compression to encode entire action chunks as a small number of discrete tokens, allowing a standard autoregressive VLM head to play the role of a robot policy without diffusion or flow-matching sampling. PI publishes π-0-FAST checkpoints and tokenizer code via the openpi GitHub repository alongside the flow-matching π-0 family, giving researchers both autoregressive and flow-matching policy baselines from the same backbone and data.

Arhitektūra

Autoregressive Vision-Language-Action policy with FAST action tokenizer

π-0-FAST keeps the PaliGemma 3B backbone (Gemma LLM + SigLIP vision tower) from π-0 but replaces the flow-matching action expert with an autoregressive decoder that emits actions as FAST tokens. FAST encodes a chunk of continuous actions by applying a Discrete Cosine Transform along the time axis and quantising the resulting frequency coefficients, yielding a compact discrete representation that captures both fast and slow motion components efficiently. The VLM is then trained with a standard next-token objective to predict these action tokens given image observations, proprioception and a natural-language instruction. This makes π-0-FAST architecturally similar to OpenVLA / RT-2-X (token-output VLA) but with a much more sample-efficient action codebook. Reported results show π-0-FAST matching or outperforming the flow-matching π-0 on many benchmarks while simplifying inference to a single autoregressive forward pass per action chunk.

Parametri

~3B (PaliGemma backbone with FAST action head)

Iespējas

• Autoregressive VLA variant of π-0 using FAST action tokens
• FAST tokenizer compresses action chunks via DCT
• Single set of weights for many robot embodiments
• Same PaliGemma 3B backbone as π-0 and π-0.5
• Matches or exceeds flow-matching π-0 on key benchmarks
• Easier integration with standard LLM serving stacks
• Open-source code and weights via openpi repository
• Compatible with existing autoregressive fine-tuning recipes
• Best for: research on autoregressive VLAs and action tokenisation.

Apmācība un licence

Trained on the same multi-embodiment teleoperation corpus used for π-0 (~10,000+ hours), plus Open-X-Embodiment data, with the FAST tokenizer providing the discrete action target instead of continuous flow-matching trajectories.

Licence: Partially open-source via the openpi GitHub repository; research weights and FAST tokenizer published, commercial deployment governed by Physical Intelligence directly.

Drošības pārbaudes: No formal RSP. Safety in deployment relies on hardware-level constraints, supervised piloting, and downstream controllers - PI's standard approach for its VLA family.