# OpenPI

> Source: https://aiwiki.ai/wiki/openpi
> Updated: 2026-07-13
> Categories: AI Models, Developer Tools, Open Source AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**OpenPI** (stylized `openpi`) is the open-source repository of robot foundation models, training code, and inference utilities published by [Physical Intelligence](/wiki/physical_intelligence), the San Francisco robotics and AI startup co-founded in early 2024 by Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, and Lachy Groom. Hosted at `github.com/Physical-Intelligence/openpi` and released under the [Apache 2.0 license](/wiki/apache_2_0_license), the repository packages pre-trained checkpoints for the company's [π0](/wiki/pi0), π0-FAST, and [π0.5](/wiki/pi_0_5) [vision-language-action models](/wiki/vision_language_action_model) (VLAs) together with [JAX](/wiki/jax) and [PyTorch](/wiki/pytorch) implementations, fine-tuning pipelines, dataset converters, a websocket-based policy server, and worked examples for common research platforms such as ALOHA, DROID, LIBERO, and the UR5 arm.[1][2]

The initial public release on February 4, 2025, marked an unusual step for a venture-backed robotics company.[1] Rather than keeping its frontier checkpoints proprietary, Physical Intelligence published the weights of its first generalist policy alongside the autoregressive π0-FAST variant, framing the move as a contribution to community research and a way to accelerate the broader goal of "general-purpose physical intelligence."[1] The release has been frequently compared to Meta's [LLaMA](/wiki/llama) strategy in language modeling,[16] and within about a year openpi had become the de facto reference implementation for generalist robot policies, drawing more than 12,000 GitHub stars and over 2,000 forks and direct adoption inside Hugging Face's [LeRobot](/wiki/lerobot) library.[2]

## Background

Physical Intelligence was founded in March 2024 with the explicit mission of building a single neural network that can control any robot to perform any physical task. The team brought together academic robotics leaders, including Stanford professor Chelsea Finn and UC Berkeley professor Sergey Levine, with practitioners from Google Brain Robotics such as CEO Karol Hausman and Brian Ichter. Within seven months the company announced π0, its first generalist policy, in a paper posted to arXiv on October 31, 2024.[3] The paper described π0 as "a novel flow matching architecture built on top of a pre-trained vision-language model (VLM) to inherit Internet-scale semantic knowledge."[3] π0 was demonstrated folding laundry, bussing a table, assembling cardboard boxes, and packing grocery bags using a dual-arm mobile manipulator, and it could be fine-tuned to a variety of downstream tasks from one to twenty hours of additional teleoperated data.[3]

At the time of the π0 announcement, the model was closed. Physical Intelligence published evaluation videos and the architectural recipe but did not release weights. The `openpi` repository was created on October 21, 2024, and remained private for several months while the team prepared a reproducible code release and coordinated a parallel PyTorch port with Hugging Face.[7] The repository was made public on February 4, 2025, the same day the company published its "Open Sourcing π0" blog post.[1] In that post, Physical Intelligence said its aim was "to enable anyone to experiment with fine-tuning π0 to their own robots and tasks," adding that by making the model available to everyone "we hope to contribute to progress toward broadly capable and general-purpose physical intelligence."[1] The release came about three months after the company closed a $400 million Series A in November 2024 at a $2.4 billion valuation, a round led by Jeff Bezos, Thrive Capital, and Lux Capital with participation from OpenAI, Redpoint Ventures, and Bond, and roughly eight months after a $70 million seed round in March 2024.[25] Publishing frontier checkpoints so soon after raising at a multibillion-dollar valuation made the open-source decision a deliberate strategic statement rather than a routine engineering deliverable.

The release positioned openpi alongside a small group of open generalist robot policies, including [OpenVLA](/wiki/openvla) from Stanford and the [RT-1 and RT-2](/wiki/open_x_embodiment) lineage at Google. The openpi release was the first time a flow-matching VLA pre-trained on a private corpus of approximately ten thousand hours of robot teleoperation data was placed in the public domain, and it was the first generalist policy from a well-funded industrial lab to ship with both inference and fine-tuning support out of the box.[1][2]

Physical Intelligence continued to raise capital aggressively while keeping its newest internal models closed. On November 20, 2025, Bloomberg reported that the company had closed a roughly $600 million Series B round at a post-money valuation of about $5.6 billion, led by CapitalG, Alphabet's independent growth fund, with participation from existing backers including Lux Capital and Thrive Capital, Jeff Bezos, and new investors Index Ventures and T. Rowe Price.[17][18] The round more than doubled the $2.4 billion valuation the company had carried since its November 2024 Series A and brought its total funding to roughly $1.1 billion.[18] In late March 2026, Bloomberg and TechCrunch reported that Physical Intelligence was in talks to raise about another $1 billion at a valuation exceeding $11 billion, with Founders Fund set to participate and Lightspeed Venture Partners in discussions alongside returning investors; the reports described the round as not yet closed and subject to change.[19][20] None of these later rounds changed the open-source status of the published openpi checkpoints, which continued to lag the company's internal frontier.

## Release timeline

The openpi repository has expanded steadily since the initial public push, with new model checkpoints, new framework support, and incremental improvements to the DROID training recipe. The table below summarizes the major milestones documented in the repository's `Updates` section and in Physical Intelligence's research blog.

| Date | Milestone | Details |
| --- | --- | --- |
| October 21, 2024 | Repository created | Private placeholder under the Physical-Intelligence GitHub organization |
| October 31, 2024 | π0 paper posted | "π0: A Vision-Language-Action Flow Model for General Robot Control" appears on arXiv |
| February 4, 2025 | Public release | Apache-licensed openpi made public with π0, π0-FAST, and seven fine-tuned checkpoints. Hugging Face publishes a parallel PyTorch port. |
| April 22, 2025 | π0.5 paper | "π0.5: a Vision-Language-Action Model with Open-World Generalization" posted to arXiv (closed weights at the time) |
| June 2025 | DROID training recipe | Repository ships instructions for training pi0-FAST-DROID style models on the full public DROID dataset |
| September 2025 | π0.5 weights released | Base π0.5 checkpoint added to openpi, along with π0.5-LIBERO and π0.5-DROID variants |
| September 2025 | Native PyTorch support | Repository ships first-class PyTorch implementations of π0 and π0.5 alongside the JAX originals |
| September 2025 | Improved DROID idle filter | Data filtering improvements added for DROID-based training runs |
| November 17, 2025 | π*0.6 paper | "π*0.6: a VLA That Learns From Experience" published, introducing the RECAP reinforcement-learning recipe. Weights are not added to the openpi repository.[5][21] |
| April 16, 2026 | π0.7 announced | "π0.7: a Steerable Model with Emergent Capabilities" published, a single generalist model that matches fine-tuned specialists. Weights are not added to the openpi repository.[22][23] |
| 2026 (ongoing) | Maintenance and tooling | 2026 commits focus on infrastructure rather than new checkpoints, for example gsutil-based downloads for Google Cloud Storage URLs, configurable PyTorch compilation modes, and race-condition fixes in the asset downloader.[2] |

### 2025-2026 developments

Through the first half of 2026 the openpi repository remained actively maintained but received no new flagship checkpoints, even as Physical Intelligence published a steady stream of internal research that moved well beyond the open π0.5 line. By mid-2026 the repository had grown to about 12,800 stars and roughly 2,200 forks, up from around 11,000 stars a year after its February 2025 launch, along with a large backlog of open issues and pull requests.[2][16] The most recent code changes through the first half of 2026 were maintenance and tooling rather than model releases: examples include switching public-asset downloads to `gsutil` for Google Cloud Storage URLs, allowing the PyTorch compilation mode to be configured, fixing a hardcoded action dimension in the π0 PyTorch model, and resolving a race condition in the `maybe_download` helper.[2] Community contributors opened pull requests to add a PyTorch implementation of LoRA fine-tuning and LoRA-aware JAX-to-PyTorch checkpoint conversion, addressing two of the known gaps in the PyTorch training path, though these remained under review rather than merged into the main branch as of mid-2026.[2]

The widening gap was driven by the research cadence on Physical Intelligence's side. The company's November 2025 "π*0.6" paper, titled "π*0.6: a VLA That Learns From Experience," introduced RECAP, short for RL with Experience and Corrections via Advantage-conditioned Policies, and reported that on its hardest tasks the recipe "more than doubles task throughput and roughly halves the task failure rate."[5] Humanoids Daily reported that the company framed the result as a sign that "RL is back" in robot learning, after a period in which most leading labs had bet on imitation learning, and independent write-ups such as Federico Sarrocco's walkthrough traced how RECAP teaches robots to learn from their own mistakes.[14][15] After that paper, the company posted "Emergence of Human to Robot Transfer in VLAs" on December 16, 2025, and "Moravec's Paradox and the Robot Olympics" on December 22, 2025.[21] In 2026 it published "The Physical Intelligence Layer" on February 24, "VLAs with Long and Short-Term Memory" on March 3 (introducing a Multi-Scale Embodied Memory, or MEM, that gives a policy both long-term and short-term memory), and "Precise Manipulation with Efficient Online RL" on March 19.[21] None of these results shipped weights to openpi, so the openpi-distributed line-up continued to stop at π0.5 while the company's internal models advanced through the π0.6 and π0.7 generations.

#### π0.7

On April 16, 2026, Physical Intelligence announced π0.7 in a blog post and accompanying paper titled "π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities."[22][23] The company described π0.7 as a single unified model that exhibits a step-change in generalization, performing dexterous tasks across robots, scenes, and skills with out-of-the-box performance the company says matches its earlier fine-tuned specialist models. The blog post said π0.7 "exhibits the first signs of compositional generalization, recombining skills from various tasks to solve new problems."[22] According to the paper, π0.7 is a roughly 5-billion-parameter VLA built from a roughly 4-billion-parameter VLM backbone initialized from Google's Gemma 3, a MEM-style video-history encoder, and an action expert of about 860 million parameters, trained with diverse multimodal prompts that include language, metadata, control modalities, and visual subgoal images describing both what to do and how to do it.[23]

Compositional generalization, the ability to recombine skills learned in different contexts to solve tasks never seen in training, is the headline claim. Physical Intelligence demonstrated it by having a robot load a sweet potato into an air fryer it had no direct training data for, succeeding when a human walked it through the steps with verbal coaching, and by having the model fold laundry on a bimanual UR5e system for which no laundry-folding data had been collected.[22] The company reported that the same π0.7 model matched or exceeded the throughput of its RL-trained π*0.6 specialists on tasks such as espresso making, box folding, and laundry folding, while cautioning that standardized robotics benchmarks largely do not exist, which makes external validation difficult.[22] As with π*0.6, the π0.7 weights and training code were not released, so the model is documented in Physical Intelligence's research output but is not part of the open openpi distribution.[2][22]

## Which models does openpi include?

The openpi repository distributes checkpoints in two tiers. *Base models* are pre-trained on Physical Intelligence's private corpus of approximately ten thousand hours of robot data drawn from seven distinct robot platforms and from the [Open X-Embodiment](/wiki/open_x_embodiment) collection, and they are intended as starting points for downstream fine-tuning.[1][2] *Fine-tuned expert checkpoints* are derived from those base models and shipped with the explicit goal of running directly on a specific robot for a specific task family. All checkpoints are hosted in the public Google Cloud Storage bucket `gs://openpi-assets/checkpoints/` and are downloaded on demand into `~/.cache/openpi` by the `openpi.shared.download.maybe_download` helper.[2]

### Base checkpoints

The three base models share the same overall recipe of a frozen-or-co-trained [PaliGemma](/wiki/paligemma) vision-language backbone augmented with a smaller action expert, but they differ in the action head and the generalization recipe.

| Checkpoint | Approximate parameters | Action head | Training data | Use case |
| --- | --- | --- | --- | --- |
| π0 base | About 3.3 billion (3 billion PaliGemma backbone plus ~300 million action expert) | Flow matching, continuous actions | ~10,000 hours of internal robot data plus Open X-Embodiment | Fine-tuning starting point for dexterous manipulation |
| π0-FAST base | About 3.3 billion | Autoregressive, discrete tokens via the FAST action tokenizer | Same as π0 base | Fine-tuning when language following and faster training matter more than the smoothest possible motion |
| π0.5 base | About 3.3 billion | Flow matching with knowledge insulation between the VLM and action expert during pre-training | Internal robot data plus large web image-text corpora and verbal task descriptions | Open-world generalization to new homes and new objects |

π0.5 is the open-world generalization successor to π0. Its paper opens by noting that "in order for robots to be useful, they must perform practically relevant tasks in the real world, outside of the lab," and it demonstrates a single policy cleaning entire kitchens and bedrooms in homes it never encountered during training.[4] The π0.5 base checkpoint formalizes a technique Physical Intelligence calls knowledge insulation, which fine-tunes the VLM backbone on FAST-tokenized discrete actions for fast representation learning while simultaneously adapting the continuous-action expert without letting its gradients propagate back into the backbone, so the policy keeps its web-scale knowledge while gaining precise motor control.[4][26]

Physical Intelligence has indicated that the π0.6 line introduced in the November 2025 "π*0.6" paper builds on the π0.5 architecture with the addition of RECAP (RL with Experience and Corrections via Advantage-conditioned Policies), a [reinforcement learning](/wiki/reinforcement_learning) recipe that uses an advantage-conditioned classifier head to refine a base VLA from autonomous practice data.[5] As of mid-2026 the π*0.6 and the newer π0.7 weights are documented in the research papers and discussed in openpi GitHub issues but have not been published in the openpi repository, so the openpi-distributed line-up currently stops at π0.5.[2][21] The closest open replication of the π0.6 line is a community effort rather than an official release: EXLA AI's `exla-ai/openpie-0.6` model card on Hugging Face describes a fully open-source PyTorch reimplementation of π0.6 trained with RECAP, a roughly 7.2-billion-parameter model split into a 5.9-billion-parameter policy and a 1.3-billion-parameter value function, distributed under Apache 2.0 and reporting metrics such as a value-function correlation of 0.986 and an action mean-squared error of 0.010, benchmarked against the figures quoted in the π*0.6 paper.[24]

### Fine-tuned expert checkpoints

The initial public release shipped a row of "expert" checkpoints tuned for widely-available research platforms, and additional rows have been added over time as new base models arrived.

| Checkpoint | Base model | Robot platform | Capability |
| --- | --- | --- | --- |
| π0-FAST-DROID | π0-FAST | DROID Franka single-arm cell | Zero-shot table-top manipulation in new scenes following language commands |
| π0-DROID | π0 | DROID Franka single-arm cell | Faster inference than π0-FAST-DROID with weaker language following |
| π0-ALOHA-towel | π0 | ALOHA dual-arm | Folds diverse towels zero-shot |
| π0-ALOHA-tupperware | π0 | ALOHA dual-arm | Unpacks food from a tupperware container |
| π0-ALOHA-pen-uncap | π0 | ALOHA dual-arm | Uncaps a pen, trained on public ALOHA data |
| π0.5-LIBERO | π0.5 | LIBERO simulation benchmark | State-of-the-art LIBERO scores |
| π0.5-DROID | π0.5 | DROID Franka single-arm cell | Fast inference plus strong language following, trained with knowledge insulation |

Many of these checkpoints have since been re-uploaded by Hugging Face into the `lerobot` namespace, where they can be loaded directly with the standard LeRobot policy API.[7] The Hugging Face mirror does not change the underlying weights; it repackages them in the format that LeRobot uses for distribution.

## What is inside the openpi repository?

The openpi tree is organized around three roles: training, inference, and platform glue. The Python package itself lives under `src/openpi/` and is split into a JAX implementation in `src/openpi/models/` and a PyTorch implementation in `src/openpi/models_pytorch/`. Top-level scripts in `scripts/` cover training, normalization-statistic computation, checkpoint conversion, and the policy server. Reference platform integrations live under `examples/`, with subfolders for ALOHA real, ALOHA simulator, DROID, LIBERO, UR5, and a `simple_client` test harness for trying inference without a robot. A separate `openpi-client` Python package is provided so that robot-side code can talk to the policy server without pulling in the full training stack.[2]

The training entry point for JAX is `scripts/train.py`, which accepts a named config from `openpi.training.config` and reads its dataset through the LeRobot v2 data loader. Training supports model parallelism via fully sharded data parallelism (FSDP) by setting `fsdp_devices` in the config, supports LoRA fine-tuning, and supports exponential moving average (EMA) weights. The PyTorch entry point is `scripts/train_pytorch.py`, which uses `torchrun` for multi-GPU and multi-node runs but as of late 2025 does not yet support FSDP, LoRA, EMA, mixed precision, or the autoregressive π0-FAST model. The PyTorch path is validated end-to-end on LIBERO and offers full feature parity for π0 and π0.5 inference. Through 2026, community contributors opened pull requests aimed at narrowing this gap, including a PyTorch LoRA fine-tuning implementation and LoRA-aware checkpoint conversion, but these remained under review rather than merged into the main branch.[2]

Inference can be run in-process by calling `policy.infer(observation)` on a `TrainedPolicy` object loaded from a checkpoint directory, or remotely by spinning up `scripts/serve_policy.py`, which listens on port 8000 and streams actions over a websocket. Remote inference is the recommended pattern for real robots because it lets the heavy GPU live on a workstation while the robot runtime stays on a real-time controller or an industrial PC, with no requirement that both machines share a Python environment or CUDA version.

## What robots does openpi support?

Physical Intelligence is explicit that openpi is an experiment and that not every robot will work. The seven robot embodiments in the π0 pre-training mixture were the company's own, and porting to new platforms requires writing observation and action adapters, computing normalization statistics, and usually fine-tuning.[2] The repository ships first-class examples for the platforms below.

| Platform | Form factor | Provided example |
| --- | --- | --- |
| DROID | Franka Panda single-arm cell with two side cameras and a wrist camera | π0-FAST-DROID and π0.5-DROID zero-shot table-top manipulation, plus full-DROID training recipe |
| ALOHA (real) | Trossen Robotics ViperX dual-arm low-cost teleoperation rig | Pen-uncap, towel-folding, and tupperware checkpoints, plus fine-tuning recipe |
| ALOHA (sim) | MuJoCo simulator for ALOHA | Sanity-check fine-tuning workflow |
| LIBERO | MuJoCo simulation benchmark of 130 household tasks | π0.5-LIBERO inference plus full fine-tuning pipeline |
| UR5 | Universal Robots UR5 industrial arm | Skeleton for adapting policies to UR5-style cells |
| Custom / experimental | Any platform with a client implementation | Simple websocket client template, with the caveat that results vary by embodiment |

The community has also stood up support for additional platforms outside the main tree, including third-party ports for the Hello Robot Stretch mobile manipulator, the Unitree H1 humanoid, and Trossen's Koch low-cost arm. Some of these live as forks; others are merged through Hugging Face's LeRobot repository, which now treats π0 and π0.5 as first-class policy classes.[7][8]

## What frameworks and hardware does openpi need?

The original implementation choice was JAX, the same framework that Physical Intelligence used internally to train π0. JAX was a natural fit because PaliGemma is itself a JAX-and-Flax model under the hood and because flow-matching training benefits from JAX's `pmap` and `pjit` distribution primitives. JAX-side dependencies are managed with `uv`, the Astral package manager, and the repository tracks a known-good lockfile so that `uv sync` reproduces the exact dependency set used by Physical Intelligence's CI. Ubuntu 22.04 is the only officially tested OS, although community users routinely run the code on other Linux distributions and inside Docker on macOS.

In September 2025 the team added a second implementation track in PyTorch, in part because the broader research community had standardized on PyTorch and in part because Hugging Face had already invested in a PyTorch port for LeRobot.[7] The PyTorch port lives at `src/openpi/models_pytorch/` and ships with a converter script, `examples/convert_jax_model_to_pytorch.py`, that translates a JAX checkpoint directory into a PyTorch state dictionary. The PyTorch path requires a patched copy of the Hugging Face `transformers` library, because π0 uses an AdaRMS normalization scheme and a non-updating key-value cache that the upstream Gemma implementation does not expose. The repository ships replacement files in `src/openpi/models_pytorch/transformers_replace/` and a `cp -r` step in the setup instructions to overlay them onto the user's local `transformers` install.

The minimum hardware bar for inference is modest by VLA standards. A single consumer GPU with more than eight gigabytes of VRAM is sufficient to run any published checkpoint, while full fine-tuning requires more than seventy gigabytes and so generally calls for an 80-gigabyte A100 or H100.[2] Multi-GPU training works on a single node out of the box, while multi-node training is not yet supported by the included `train.py` and `train_pytorch.py` scripts. The table below summarizes the recommended hardware for each workflow.

| Workflow | Minimum VRAM | Example GPU |
| --- | --- | --- |
| Inference | 8 GB | NVIDIA RTX 4090 |
| Fine-tuning with LoRA (JAX only) | 22.5 GB | NVIDIA RTX 4090 |
| Full fine-tuning | 70 GB | NVIDIA A100 80 GB or H100 |
| Multi-GPU full fine-tuning | 70 GB across a node | Multi-A100 or multi-H100 server |

## How does openpi integrate with LeRobot and Hugging Face?

openpi has a tight, deliberate relationship with [LeRobot](/wiki/lerobot), the Hugging Face open robotics initiative. The repository pulls LeRobot in as a Git submodule and uses LeRobot's dataset format as the canonical training-data representation, so any dataset that already lives on the Hugging Face Hub in the LeRobot v2 layout can be fed straight into the openpi training scripts after running `scripts/compute_norm_stats.py`. The LeRobot team in turn maintains a parallel PyTorch implementation of π0 and π0.5 inside the `lerobot` package, and they re-export the openpi checkpoints under names such as `lerobot/pi0_base`, `lerobot/pi05_base`, and `lerobot/pi05_libero_finetuned` on the Hugging Face Hub.[9][10] Announcing the integration, Hugging Face wrote that "both π0 and π0-FAST, developed by Physical Intelligence, are now available in the LeRobot repository, bringing generalist robotic intelligence to the Hugging Face ecosystem."[7] From a user's perspective this means there are now two equally valid ways to run a π0.5 policy: through the openpi training and inference stack with the original JAX or new PyTorch path, or through the LeRobot policy API with Hugging Face's PyTorch port.[8]

The partnership has been mutually beneficial. openpi gives LeRobot a frontier-grade pre-trained policy that would otherwise have required a multi-million-dollar pre-training run, and LeRobot gives openpi a much larger audience of hobbyist roboticists, students, and other researchers who would never have set up a JAX environment on their own. Several other community projects, including Allen Z. Ren's `open-pi-zero` PyTorch re-implementation[11] and Phil Wang's `pi-zero-pytorch` reference codebase[12], have folded openpi-derived weights into their tutorials, deepening the model's reach into the academic and indie-research ecosystem.

## Is openpi open source and free for commercial use?

The openpi repository is licensed under the [Apache 2.0 license](/wiki/apache_2_0_license), one of the most permissive widely used open-source licenses. Apache 2.0 allows commercial use, modification, distribution, sublicensing, and private use, with the standard requirements that the license text be preserved, that changes be noted, and that any trademarks held by Physical Intelligence not be used to endorse derivative products. The repository also bundles the Gemma terms of use because π0, π0-FAST, and π0.5 incorporate weights derived from Google's PaliGemma vision-language model, and downstream users are required to honor Google's prohibited-uses policy in addition to Apache 2.0. There is no separate research-only or non-commercial restriction on the openpi weights themselves, and several robotics startups have publicly demonstrated commercial fine-tunes built on top of the base checkpoints.[2]

## How was openpi received?

The February 2025 release drew immediate attention from the robot-learning research community. Within its first weeks the repository had passed five thousand GitHub stars, and within a year it sat at more than eleven thousand stars with close to two thousand forks, reaching about 12,800 stars and 2,200 forks by mid-2026 and putting it among the most-starred robotics-specific repositories on GitHub.[2] The Robot Report covered the launch, noting that Physical Intelligence had released the code and weights for π0 as an experimental repository with checkpoints for available platforms such as ALOHA and DROID, example code to run inference on real and simulated robots, and code for fine-tuning the base model to custom tasks.[6] Hugging Face's own engineering blog and multiple robotics outlets covered the release as a deliberate Meta-style play to set the de facto industry baseline.[7] Independent researchers, including the University of Pennsylvania's PAL Lab, quickly published "in the wild" evaluation studies of π0-FAST-DROID, finding that the DROID-tuned checkpoint generalized strikingly well to novel objects and cluttered scenes while exposing failure modes such as freezing mid-task, imprecise spatial reasoning, and sensitivity to how instructions are phrased. The Penn team concluded that "π0 produces sensible behaviors across a wide variety of our tasks, although it is important to note that sensible behaviors are often insufficient for task completion."[13]

The release also triggered a wave of community re-implementations. Allen Z. Ren's `open-pi-zero` provides a from-scratch PyTorch rewrite of the π0 architecture,[11] lucidrains' `pi-zero-pytorch` offers a reference implementation aimed at teaching the flow-matching action head,[12] and the `exla-ai/openpie-0.6` model card on Hugging Face advertises a community PyTorch port of the not-yet-officially-released π0.6 line trained with RECAP.[24] By late 2025 OpenVLA, RT-2-X, Octo, and openpi together formed a recognizable family of generalist robot policies that researchers could pick from when starting a new project. Industry analysts such as TSG Invest have argued that the openpi strategy mirrors Meta's LLaMA play, using an open weight release to make the company's tooling and conventions the default substrate for downstream robotics work even when the most capable internal models remain proprietary.[16] That dynamic became more pronounced through 2026 as Physical Intelligence's internal models advanced through the π0.6 and π0.7 generations while the open openpi line-up stayed at π0.5.[21][22]

## What are openpi's limitations?

Physical Intelligence has been candid that openpi is a research artifact rather than a productionized robotics stack. The README opens with an explicit warning that π0 "was developed for our own robots, which differ from the widely used platforms such as ALOHA and DROID," and that the team does "not expect every such attempt to be successful," concluding: "π0 may or may not work for you, but you are welcome to try it and see!"[2] Several capabilities are missing from the open release relative to Physical Intelligence's internal stack. There is no internal pre-training data, so users who want to extend the pre-training mixture must contribute their own teleoperation logs. There is no support for multi-node training, no released checkpoint for the more recent π*0.6 or π0.7 lines, and the PyTorch implementation does not yet support FSDP, LoRA, EMA, mixed-precision training, or the π0-FAST model.[2]

Licensing is also more nuanced than the headline Apache 2.0 label suggests. The action expert weights inherit terms from PaliGemma, which carries the Gemma terms of use, so commercial deployments need to review both Apache 2.0 and the Gemma prohibited-uses policy before shipping a derived product.[2] Finally, evaluations by the University of Pennsylvania PAL Lab and other groups have shown that π0 and π0.5 can struggle with long-horizon tasks, fine-grained dexterous manipulation, and integrating verbal feedback during execution; the Penn study documented failure modes including early stopping on multi-step tasks, imprecise spatial reasoning, prompt sensitivity, and freezing when the wrist camera is occluded.[13] The π*0.6 line with its RECAP recipe and the later π0.7 model are Physical Intelligence's research directions for addressing some of these limitations through autonomous practice, reinforcement learning, and steerable multimodal prompting, but those advances are not yet reflected in the openpi distribution.[21][22]

## See also

- [π0](/wiki/pi0) - the original flow-matching vision-language-action model whose weights are distributed through openpi
- [π0.5](/wiki/pi_0_5) - the open-world generalization successor trained with knowledge insulation, also distributed through openpi
- [Physical Intelligence](/wiki/physical_intelligence) - the San Francisco startup that publishes openpi
- [LeRobot](/wiki/lerobot) - the Hugging Face open robotics library that integrates openpi checkpoints
- [OpenVLA](/wiki/openvla) - an earlier open vision-language-action model trained on Open X-Embodiment
- [Open X-Embodiment](/wiki/open_x_embodiment) - the cross-embodiment dataset used in part to pre-train openpi base models
- [Vision-language-action model](/wiki/vision_language_action_model) - the broader class of models that openpi belongs to
- [Apache 2.0 license](/wiki/apache_2_0_license) - the license that governs the openpi code and weights

## References

1. Physical Intelligence. "Open Sourcing π0." Blog post, February 4, 2025. https://www.pi.website/blog/openpi
2. Physical-Intelligence/openpi. GitHub repository README. https://github.com/Physical-Intelligence/openpi
3. Black, Kevin et al. "π0: A Vision-Language-Action Flow Model for General Robot Control." arXiv preprint, October 31, 2024. https://arxiv.org/abs/2410.24164
4. Intelligence, Physical. "π0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv preprint, April 22, 2025. https://arxiv.org/abs/2504.16054
5. Intelligence, Physical. "π*0.6: a VLA That Learns From Experience." arXiv preprint, November 18, 2025. https://arxiv.org/abs/2511.14759
6. Crowe, Steve. "Physical Intelligence open-sources Pi0 robotics foundation model." The Robot Report, February 7, 2025. https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/
7. Hugging Face. "π0 and π0-FAST: Vision-Language-Action Models for General Robot Control." Hugging Face Blog, February 4, 2025. https://huggingface.co/blog/pi0
8. Hugging Face. "π0.5 (Pi05) Policy." LeRobot documentation. https://huggingface.co/docs/lerobot/pi05
9. Hugging Face. "lerobot/pi05_base." Model card. https://huggingface.co/lerobot/pi05_base
10. Hugging Face. "lerobot/pi0_base." Model card. https://huggingface.co/lerobot/pi0_base
11. allenzren. "open-pi-zero: Re-implementation of pi0 vision-language-action (VLA) model from Physical Intelligence." GitHub repository. https://github.com/allenzren/open-pi-zero
12. lucidrains. "pi-zero-pytorch: Implementation of π0, the robotic foundation model architecture proposed by Physical Intelligence." GitHub repository. https://github.com/lucidrains/pi-zero-pytorch
13. Penn PAL Lab. "Evaluating π0 in the Wild: Strengths, Problems, and the Future of Generalist Robot Policies." Project page, 2025. https://penn-pal-lab.github.io/Pi0-Experiment-in-the-Wild/
14. Sarrocco, Federico. "π*0.6 and RECAP: Teaching Robots to Learn From Their Mistakes." Blog post, November 2025. https://federicosarrocco.com/blog/pi-star-06-recap
15. Humanoids Daily. "Physical Intelligence Claims 'RL is Back' With New Model That Learns From Its Own Mistakes." November 2025. https://www.humanoidsdaily.com/news/physical-intelligence-claims-rl-is-back-with-new-model-that-learns-from-its-own-mistakes
16. TSG Invest. "Physical Intelligence Stock: $11B Valuation - A Buy?" Analyst commentary. https://tsginvest.com/physical-intelligence/
17. Metz, Rachel, and Kate Clark. "Robotics Startup Physical Intelligence Valued at $5.6 Billion in New Funding." Bloomberg, November 20, 2025. https://www.bloomberg.com/news/articles/2025-11-20/robotics-startup-physical-intelligence-valued-at-5-6-billion-in-new-funding
18. Demaitre, Eugene. "Physical Intelligence raises $600M to advance robot foundation models." The Robot Report, November 2025. https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/
19. Bloomberg. "Ex-DeepMind Staffers' Robotics Startup in Talks for $11 Billion Valuation." March 27, 2026. https://www.bloomberg.com/news/articles/2026-03-27/ex-deepmind-staffers-robotics-startup-in-talks-for-11-billion-valuation
20. Wiggers, Kyle. "Physical Intelligence is reportedly in talks to raise $1B, again." TechCrunch, March 27, 2026. https://techcrunch.com/2026/03/27/physical-intelligence-is-reportedly-in-talks-to-raise-1-billion-again/
21. Physical Intelligence. "Blog." Research blog index. https://www.pi.website/blog
22. Physical Intelligence. "π0.7: a Steerable Model with Emergent Capabilities." Blog post, April 16, 2026. https://www.pi.website/blog/pi07
23. Physical Intelligence. "π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities." Paper, April 16, 2026. https://www.pi.website/download/pi07.pdf
24. EXLA AI. "exla-ai/openpie-0.6." Hugging Face model card. https://huggingface.co/exla-ai/openpie-0.6
25. SiliconANGLE. "AI startup Physical Intelligence raises $400M to create a brain for any robot." November 4, 2024. https://siliconangle.com/2024/11/04/ai-startup-physical-intelligence-raises-400m-create-brain-robot/
26. Physical Intelligence. "Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better." arXiv preprint, May 2025. https://arxiv.org/abs/2505.23705