OpenPI
Robotics modelVision-language-action model code, checkpoints and adaptation examples for pi0, pi0-FAST and pi0.5, with backend-specific features and robot-specific action conventions. [1]
Setup & usage →CheckpointsModel hub / homepageGetting started
- For
- Robot-learning researchers able to validate model, data and robot-interface compatibility [1]
- Parameters
- ~3B class (editorial estimate inferred from gemma_2b + gemma_300m variants in pi0_config.py [10]
- Source rights
- Some licenses are unresolved. Review licenses
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- Ecosystem
- LeRobot
- License
- Apache-2.0 · license not fully assessed for model weights, datasets, documentation · Details by part
Overview
OpenPI (Physical-Intelligence/openpi) provides training code, remote serving utilities, and downloadable checkpoints for Physical Intelligence's pi0, pi0-FAST, and pi0.5 Vision-Language-Action (VLA) models. At the pinned commit (215abfb217db), Pi0Config pairs a PaliGemma (gemma_2b) vision-language backbone with a gemma_300m flow-matching action expert (pi0.5 adds discrete state input and adaRMSNorm), while pi0-FAST uses an autoregressive FAST action tokenizer. Based on those configured variants, this profile editorially estimates a ~3B-class parameter footprint, though the exact parameter count is unverified because GCS checkpoint binaries were not downloaded. Upstream describes the release as experimental and warns that adapting a policy to another robot may not succeed. [1] [2] [10]
Model card summary
- Architecture
PaliGemma (gemma_2b) vision-language backbone paired with a Gemma-300M flow-matching action expert (pi0 and pi0.5 with adaRMSNorm) or autoregressive FAST action tokenizer (pi0-FAST).
- Base model
PaliGemma (gemma_2b) + Gemma-300M action expert
- Parameter count
~3B class (editorial estimate inferred from gemma_2b + gemma_300m variants in pi0_config.py; exact tensor count unverified)
Inferred from gemma_2b + gemma_300m variants in Pi0Config; GCS checkpoint binaries not downloaded.
[10] · Editorial assessment
- Input modalities
- rgbproprioceptionlanguage
Three RGB image slots (base, left wrist, right wrist), proprioceptive state, and language prompt.
- Output modalities
- joint-positionsjoint-velocitiesgripper-command
Base models use joint angles (radians) + gripper [0, 1]; DROID uses joint-velocity actions.
- Action space
Continuous action chunks with default action_horizon: 50 and action_dim: 32 (padded) in Pi0Config; joint position (radians) or joint velocity (DROID) and gripper conventions depend on the selected robot/data config.
- Control frequency
15 Hz documented for DROID; other embodiments depend on robot/data config
- Trained / supported embodiments
- DROID (Franka)
- ALOHA (real and sim)
- UR5
- LIBERO simulation (Franka)
- Training datasets
- 10k+ hours of robot manipulation data (mixed internal and public datasets; public fine-tuning examples include DROID and LIBERO)
Checkpoints
| Checkpoint | Format | Size | Access | Pinned revision |
|---|---|---|---|---|
| pi0_base (gs://openpi-assets/checkpoints/pi0_base) | jax-orbax | Not stated | open | 215abfb217db |
| pi0_fast_base (gs://openpi-assets/checkpoints/pi0_fast_base) | jax-orbax | Not stated | open | 215abfb217db |
| pi05_base (gs://openpi-assets/checkpoints/pi05_base) | jax-orbax | Not stated | open | 215abfb217db |
| pi05_droid (gs://openpi-assets/checkpoints/pi05_droid) | jax-orbax | Not stated | open | 215abfb217db |
| pi05_libero (gs://openpi-assets/checkpoints/pi05_libero) | jax-orbax | Not stated | open | 215abfb217db |
| pi0_fast_droid (gs://openpi-assets/checkpoints/pi0_fast_droid) | jax-orbax | Not stated | open | 215abfb217db |
Checkpoints are hosted in gs://openpi-assets/checkpoints/ and can be converted to PyTorch for pi0/pi0.5; payloads not downloaded so size_gb is null.
Is this for you?
- Use cases
- Fine-tuning pi0, pi0-FAST, or pi0.5 base checkpoints on LeRobot-converted datasets
- Remote WebSocket GPU policy serving to DROID, ALOHA, UR5, or LIBERO setups
- Known limitations
- PyTorch lacks pi0-FAST, LoRA, FSDP, mixed-precision training and EMA training weights in this snapshot
- pi0.5 implementation supports the flow-matching head
- Fine-tuned checkpoints may not generalize to a different robot setup
- Checkpoint and complete pretraining-data rights are not established by the code license
Rights scope is an editorial limitation; backend support follows the pinned README.
- Safety and limitations
- PyTorch backend lacks pi0-FAST, LoRA, FSDP, mixed-precision training, and EMA weights in this snapshot.
- pi0.5 implementation supports the flow-matching head only in this repository.
- Fine-tuned checkpoints may not generalize to a different robot setup; action-space and normalization conventions must match.
- Bundled Gemma Terms of Use include use restrictions; complete checkpoint and 10k+ hour pretraining-data rights are not established by the Apache-2.0 code license.
Compute & fine-tuning
- Inference compute
- Minimum VRAM: 8 GB
- Documented hardware: NVIDIA RTX 4090 (upstream >8 GB estimate)
- Precision:
bf16 - Edge / consumer capable: No
Upstream single-GPU estimate (>8 GB VRAM) in bfloat16; not independently measured.
- Fine-tuning
- Supported: Yes
- Methods: LoRA (JAX only; >22.5 GB estimate); Full fine-tune (JAX or PyTorch for pi0/pi0.5; >70 GB estimate)
- Minimum VRAM: 22.5 GB
- Recipe: Upstream fine-tuning documentation
PyTorch backend lacks pi0-FAST, LoRA, FSDP, mixed-precision training, and EMA weights in this snapshot.
- Host computer
NVIDIA GPU; upstream single-GPU estimates exceed 8 GB for inference, 22.5 GB for LoRA fine-tuning and 70 GB for full fine-tuning.
Upstream estimates, not measurements here; LoRA support is backend-dependent.
- Operating system
- Ubuntu 22.04
Upstream states tested and supported platform for this snapshot.
- Dependencies
- uv-managed Python environment and repository submodules
- JAX or supported PyTorch model implementation
- Matching checkpoint, normalization statistics and policy configuration
- Robot-specific observation/action transformations
Setup & usage
Open the upstream getting-started guide [1] · Model card / documentation [4] [6]
- Clone recursively with submodules and install environment via uv on Ubuntu 22.04
- Choose JAX or PyTorch backend (note PyTorch patches transformers 4.53.2)
- Download or convert a matched checkpoint from gs://openpi-assets/checkpoints/ and compute or reuse normalization statistics
- Validate policy I/O with the no-robot client or serve_policy.py before hardware integration
Upstream-reported benchmarks
Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.
| Benchmark | Upstream-reported score | Scope & caveats |
|---|---|---|
| LIBERO (Spatial, Object, Goal, LIBERO-10) | 96.85% average (Spatial 98.8, Object 98.2, Goal 98.0, LIBERO-10 92.4) for pi0.5 @ 30k finetuned | Upstream-reported in examples/libero/README.md for pi05_libero; not independently verified by OSR. |
Project notes
Action space, checkpoints, and normalization statistics
Pi0Config defines three RGB camera slots (base_0_rgb, left_wrist_0_rgb, right_wrist_0_rgb), a proprioceptive state vector, and a tokenized language prompt, outputting continuous action chunks with default action_horizon: 50 and padded action_dim: 32. Base and fine-tuned checkpoints (pi0_base, pi0_fast_base, pi05_base, pi05_droid, pi05_libero, pi0_fast_droid, and ALOHA task variants) are hosted in gs://openpi-assets/checkpoints/ and cached on demand. Selecting a checkpoint requires matching both the policy configuration and normalization statistics: base-model conventions use joint angles in radians and gripper values from 0 (open) to 1 (closed), whereas DROID uses joint-velocity actions at 15 Hz. As a concrete next step before adaptation or hardware integration, reconcile your robot's camera keys, state ordering, and action dimensions through the selected policy's input/output transforms, then verify the contract offline with the documented no-robot inference client. [1] [5] [6] [10]
Backend differences, compute estimates, and upstream-reported benchmarks
The repository supports JAX and PyTorch, but PyTorch excludes pi0-FAST, LoRA, FSDP, mixed-precision training, and EMA weights in this snapshot, and requires patching transformers 4.53.2 (which upstream warns can propagate through uv's default hardlink cache). Upstream lists Ubuntu 22.04 with an NVIDIA GPU and single-GPU estimates of >8 GB VRAM for inference (e.g., RTX 4090, bfloat16), >22.5 GB for LoRA fine-tuning, and >70 GB for full fine-tuning, plus a WebSocket policy server (serve_policy.py) and openpi-client. In examples/libero/README.md, upstream reports a 96.85% average success rate on LIBERO (Spatial 98.8, Object 98.2, Goal 98.0, LIBERO-10 92.4) for pi05_libero fine-tuned for 30k steps. All benchmark scores and memory estimates are upstream-reported; Open Source Robotics has not downloaded checkpoints or run training, fine-tuning, inference, or robot evaluations. [1] [6] [11]
What we checked
Source research only; no installation, physical build or performance test performed. No affiliate links added.
Updates
- Upstream last updated
- Aug 24, 2026 [2]
- Latest release
- Not established
- Profile updated
- Nov 21, 2026
- Sources last checked
- Nov 21, 2026
- Next review
- Feb 19, 2027
- Reviewed revision
- 215abfb217dbac7d5f1273282331b9b1866c0479 (commit snapshot; no published stable release found) · software repository
Separate license matrix
Available project sources support research coverage; complete build and commercial qualification remain separate.
| Layer | Availability / license |
|---|---|
| Code | available · Apache-2.0 Code license; repository also includes separate Gemma terms, and dependencies retain their own terms. [3] |
| Model weights | available · License not fully assessed README publishes gs://openpi-assets checkpoint paths. Payloads not downloaded. Complete checkpoint license mapping unresolved; bundled Gemma terms are not Apache-2.0 and include use restrictions. [1] [8] |
| Training datasets | available · License not fully assessed Public fine-tuning data links and conversion examples exist. Full 10k+ hour pretraining mixture availability and constituent rights not established; includes references to internal data. [1] |
| Documentation | available · License not fully assessed Documentation source available; separate documentation-wide rights not established. [4] [5] [6] |
Separate code, weights, and pretraining-data rights
The repository source code carries an Apache-2.0 license, while the repository also bundles Gemma Terms of Use (LICENSE_GEMMA.txt) containing use restrictions and distribution obligations. Because the inspected sources do not establish a uniform license across every GCS checkpoint, this profile does not label the model weights Apache-2.0. Similarly, public fine-tuning examples (such as DROID and LIBERO) do not establish public availability or uniform licensing for the full 10k+ hour pretraining mixture, which references internal task data. [1] [3] [8]
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
Project sources
- S001 Physical-Intelligence/openpi: README.md
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026
Used for: Overview, Model card summary, Checkpoints, Is this for you? and 5 more - S002 Physical-Intelligence/openpi: commit 215abfb
- S003 Physical-Intelligence/openpi: LICENSE
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026
Used for: Is this for you?, Separate license matrix - S004 Physical-Intelligence/openpi: files
Revision, date & where cited
github.com · revision
215abfb· archive file inventory · Accessed Oct 1, 2026
Used for: Setup & usage, Separate license matrix - S005 Pinned openpi: docs/norm_stats.md
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026
Used for: Model card summary, Is this for you?, Compute & fine-tuning, Setup & usage and 2 more - S006 Pinned openpi: docs/remote_inference.md
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026
Used for: Model card summary, Is this for you?, Compute & fine-tuning, Setup & usage and 3 more - S007 Pinned openpi: pyproject.toml
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026 - S008 Bundled Gemma Terms of Use
Revision, date & where cited
github.com · revision
215abfb· Accessed Oct 1, 2026
Used for: Is this for you?, Separate license matrix - S010 Pinned openpi: src/openpi/models/pi0_config.py
Pi0Config defaults: paligemma_variant='gemma_2b', action_expert_variant='gemma_300m', dtype='bfloat16', action_dim=32, action_horizon=50, 3 RGB cameras (base_0_rgb, left_wrist_0_rgb, right_wrist_0_rgb), state vector, and tokenized prompt; pi0.5 uses discrete state input and adaRMSNorm.Revision, date & where cited
github.com · revision
215abfb· Accessed Nov 21, 2026
Used for: Overview, Model card summary, Action space, checkpoints, and normalization statistics - S011 Pinned openpi: examples/libero/README.md
Upstream-reported LIBERO benchmark table for pi0.5 @ 30k finetuned (pi05_libero): Spatial 98.8, Object 98.2, Goal 98.0, LIBERO-10 92.4, Average 96.85.Revision, date & where cited
github.com · revision
215abfb· Accessed Nov 21, 2026
Used for: Model card summary, Upstream-reported benchmarks, Backend differences, compute estimates, and upstream-reported benchmarks
Videos
- S009 Physical Intelligence (YouTube channel)
The project’s own YouTube channel, found by searching YouTube and checked against the channel page. Not reviewed by us for accuracy.Revision, date & where cited
YouTube channel Physical Intelligence · youtube.com · Accessed Oct 4, 2026
Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1
Change history
2026-10-04 — Added 1 YouTube reference(s) (S009): official channel or videos linked from the pinned upstream sources. Existing claims, source-check dates and evidence are unchanged.
2026-11-21 — Migrated 0.6.0-draft.1 -> 0.7.0-draft.5 -> 0.8.0-draft.1 (kind=model); added Pi0Config (S010) and LIBERO benchmark (S011) facts while preserving historical source dates.