π0-FAST (FAST 토크나이저)

Physical Intelligence, UC Berkeley, Stanford · 2025-01

VLAπ family3Bdiscrete tokens

Compresses action chunks with DCT and BPE so high-frequency continuous control can be learned as autoregressive tokens.

Paper

Backbone LLM: Gemma 2B · PaliGemma 3B config
Vision encoder: SigLIP 400M
Action head: FAST token autoregressive (vocabulary 1024) (discrete tokens)
Chunk: 1 s (DROID 15 steps)
Control rate: 5–50 Hz (per task)
Architecture (original summary): PaliGemma-3B, no action expert; autoregressive decoding of FAST discrete action tokens (DCT+BPE, rounding 10, vocabulary 1024), 1-second chunks (DROID 15 steps), evaluated at 5~50 Hz

Pretraining accelerators: Generalist model Not disclosed; DROID-only policy (from PaliGemma) H100×8
Pretraining duration: Generalist model Not disclosed (5× fewer GPU-hours than π0); DROID-only policy about 4 days
Pretraining steps & batch: Generalist model Not disclosed; DROID-only policy 240k iter·batch 256 (3 epochs)
GPU/TPU-hours: Not disclosed
Fine-tuning: openpi pi0_fast_full_droid_finetune: H100×8 about 2 days, 100k steps·batch 256; pi0_fast_libero 30k steps (same for LoRA version). LeRobot LIBERO: H100×8, 40k steps·batch 256 (secondary)
Data type: Real-robot teleoperation
Data scale: Robot data 10k h; FAST+ tokenizer about 1M 1-second chunks (main text)/1M trajectories (abstract); DROID 75k episodes·21M samples; LIBERO 270k samples
Main datasets: PI's own data (903M timesteps), BRIDGE v2, DROID, OXE (9.1% of mix)
Tasks & evaluation: Table bussing (UR5, 20 Hz), T-shirt folding (bimanual ARX, 50 Hz), Grocery bagging, Toast (bimanual Trossen, 50 Hz), Laundry folding; DROID zero-shot 16 tasks·44 trials (3 campuses); openpi LIBERO 96.4/96.8/88.6/60.2, average 85.5
Robot: UR5/UR5e, bimanual ARX, bimanual Trossen, DROID Franka; FAST+ training: ARX, AgileX, Fibocom Mobile, Franka FR3, UR5, Trossen, Mobile Trossen

Inference device & speed: RTX 4090 about 750 ms per chunk (decoding 30~60 action tokens); π0 under 100 ms on the same GPU
Inference memory: > 8 GB (common to openpi README)
Weights bf16 (calc.): 6 GB
Where it runs: Local PC

Reinforcement learning: Not used · FAST token imitation learning
Base control: Fixed · No base control (evaluation on fixed manipulators only; FAST+ training data includes mobile)

Key features: Compresses action chunks with DCT and BPE so high-frequency continuous control can be learned as autoregressive tokens. Cuts training GPU-hours 5× vs π0, but chunk inference is slow at about 750 ms.
Notes: Compression ratio BridgeV2 1.75×, DROID 3.6×, Bussing 5.0×, T-shirt 13.2×. FAST+ tokenizer released on HF under Apache-2.0. openpi pi0_fast_droid default horizon is 10, unlike the paper's 15 steps for DROID. 'About 5 days' in openpi README_train (240k iter from PaliGemma) refers to pi05_full_droid_finetune.
Verification: Corrections applied