π0-FAST
Autoregressive variant of π0 from Physical Intelligence: the PaliGemma VLM predicts discrete FAST action tokens (DCT + BPE compression of action chunks) that are decoded back to continuous actions.
open weightsverified 2026-10-24paper →
Architecture
Compute requirements
| Memory | Precision | Conditions | |
|---|---|---|---|
| train | ≥24 GB · ≤40 GBsrc | — | Peak training VRAM at batch size 8 with AdamW (LeRobot default), "Large VLA" group (pi0, pi0_fast, pi05, xvla, wall_x) in the LeRobot compute hardware guide; suggested A100 40 GB+, "24 GB tight at BS 1". The guide calls its figures indicative, order-of-magnitude numbers, not exact predictions. Precision not stated. · batch 8 |
| inference | ≥8 GBsrc | — | openpi README single-GPU requirement for inference ("> 8 GB", example RTX 4090), stated for all openpi models and the openpi runtime, not LeRobot. Precision not stated. No LeRobot measurement for π0-FAST found. |
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running π0-FAST 9 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| ALOHA (stationary) | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Evidence is Physical Intelligence's own zero-shot evaluation of its generalist π0-FAST on an in-house bimanual Trossen ViperX setup that mirrors ALOHA; openpi ships no π0-FAST ALOHA config or checkpoint. |
| Bimanual SO-100/SO-101 | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Franka Research 3 | works with caveats | T1 · upstream documented | 2026-10-25 today | Community-reported only (third-party EmbodiedForge/droid-fr3, a DROID fork that replaces polymetis with the Franky SDK); not documented by Physical Intelligence or maintainer-tested. |
| Franka Emika Panda | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Documented for the DROID platform (Panda, two ZED 2 exterior cameras plus a ZED Mini wrist camera, DROID control laptop/NUC), not for an arbitrary Panda setup. openpi warns that its expert checkpoints may not generalize to other setups. |
| Koch v1.1 | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| SO-100 | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| SO-101 | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Unitree G1 | partial | T3 · community reported | 2026-10-25 today | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Universal Robots UR5e | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | The evidence is the authors' own UR5e evaluation; openpi releases no UR5e π0-FAST checkpoint or UR5e training config. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, EarthRover Mini Plus, HopeJR, I2RT YAM, LeKiwi, ROBOTIS OMX, OpenArm, Reachy 2, reBot Arm B601-DM, Trossen WidowX AI, UFACTORY xArm, XLeRobot.
Framework support
| Framework | Support | Since | Type string | Docs |
|---|---|---|---|---|
| LeRobot | nativesrc | 0.4.3 | pi0_fast | docs |
| openpi | upstream-reposrc | — | — | docs |
| LeRobot: Current implementation added in v0.4.3 ("add autoregressive VLAs with tokenization PiFast"). An earlier implementation registered as `pi0fast` existed in v0.3.2–v0.3.3 and was deprecated/removed in v0.4.0 ("deprecate pi0fast"); the directory is absent at v0.4.0–v0.4.2. Default FAST tokenizer: lerobot/fast-action-tokenizer. | ||||
| openpi: Official release; base checkpoint gs://openpi-assets/checkpoints/pi0_fast_base, plus π0-FAST-DROID. | ||||
Checkpoints
- lerobot/pi0fast-base · LeRobot-format base model; Hub license tag gemma.
- lerobot/pi0fast-libero · LIBERO fine-tune used for the docs' results (82.5% average); no license tag on the Hub.
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot π0-FAST docs; PaliGemma (SigLIP + Gemma 2B) predicting discrete FAST tokens decoded to continuous actions.
“The model takes images, text instructions, and robot state as input, and outputs discrete FAST tokens that are decoded back to continuous actions.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/pi0fast.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `pi0_fast` (also registered at v0.4.3); chunk_size default 50.
“@PreTrainedConfig.register_subclass("pi0_fast")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/pi0_fast/configuration_pi0_fast.py - T1 · upstream documentedrelease-notessource 2026-01-22verified 2026-10-24 (1d ago)lerobot 0.4.3
v0.4.3 added the current π0-FAST (pi0_fast) implementation.
“* feat(policies): add autoregressive VLAs with tokenization PiFast by @jadechoghari in https://github.com/huggingface/lerobot/pull/2734”
https://github.com/huggingface/lerobot/releases/tag/v0.4.3 - T1 · upstream documentedrelease-notessource 2025-10-23verified 2026-10-24 (1d ago)lerobot 0.4.0
The earlier `pi0fast` implementation (v0.3.x) was deprecated in v0.4.0.
“* chore(policies): deprecate pi0fast by @imstevenpmwork in https://github.com/huggingface/lerobot/pull/2203”
https://github.com/huggingface/lerobot/releases/tag/v0.4.0 - T1 · upstream documentedpapersource 2025-01-16verified 2026-10-24 (1d ago)
FAST paper introduces the DCT-based action tokenizer used with the π0 VLA.
“Our tokenization approach, Frequency-space Action Sequence Tokenization (FAST), enables us to train autoregressive VLAs for highly dexterous and high-frequency tasks where standard discretization methods fail completely.”
https://arxiv.org/abs/2501.09747 - T1 · upstream documentedofficial-docsverified 2026-10-24 (1d ago)openpi 215abfb
openpi (official PI repo) ships π0-FAST with base checkpoints.
“- the [π₀-FAST model](https://www.physicalintelligence.company/research/fast), an autoregressive VLA, based on the FAST action tokenizer.”
https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/README.md - T1 · upstream documentedofficial-docsverified 2026-10-24 (1d ago)openpi 215abfb
openpi single-GPU memory requirements: inference >8 GB, LoRA fine-tuning >22.5 GB, full fine-tuning >70 GB.
“| Inference | > 8 GB | RTX 4090 |”
https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/README.md - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Compute hardware guide: Large VLA group (incl. `pi0_fast`) ~24–40GB at BS 8 with AdamW; indicative figures.
“| Large VLA | `pi0`, `pi0_fast`, `pi05`, `xvla`, `wall_x` | ~24–40GB | A100 40 GB+ (24 GB tight at BS 1) |”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/hardware_guide.mdx - T1 · upstream documentedmodel-cardsource 2026-01-09verified 2026-10-24 (1d ago)
lerobot/pi0fast-base card lists inputs, next-token objective and FAST-token action representation; license gemma.
“- **Inputs:** images (multi-view), proprio/state, optional language instruction”
https://huggingface.co/lerobot/pi0fast-base