Physical Intelligence (π0.5): The Open Reference Point for Generalist Robot Policies
What Physical Intelligence has actually demonstrated with π0.5, what stays a company claim, and why the open weights lag the flagship.
Physical Intelligence describes itself in one line on its homepage: "bringing general-purpose AI into the physical world," built by "engineers, scientists, roboticists, and company builders developing learning algorithms to create a model that will control any robot to do any task." That is the pitch. The measured record underneath it is narrower and more interesting: one vision-language-action model, π0.5, with a documented ablation table, a published scaling curve, released weights, an Apache 2.0 code repository, and a reproducible simulation benchmark. Everything else on the site is video and self-report.
This profile treats π0.5 as the anchor because it is the last model the lab opened. The company has shipped a steady cadence since October 2024: π0 (31 October 2024), FAST (16 January 2025), openpi (4 February 2025), Hi Robot (26 February 2025), π0.5 (22 April 2025), knowledge insulation (28 May 2025), real-time action chunking (9 June 2025), π*0.6 (17 November 2025), human-to-robot transfer in VLAs (16 December 2025), Moravec's paradox and the robot olympics (22 December 2025), the Physical Intelligence Layer partner post (24 February 2026), MEM (3 March 2026), precise manipulation with efficient online RL (19 March 2026), and π0.7 (16 April 2026). π0.7 is the current flagship as of August 2026. π0.5, dated 22 April 2025, sits several model generations back: π0.6, π*0.6 and π0.7 all postdate it. It is also the only recent model anyone outside the lab can run. If you are benchmarking "against Physical Intelligence" today, you are benchmarking against an April-2025 model, and a partial version of it at that.
Companion reading: foundation models for VLA robotics, imitation learning in robotics, reinforcement learning for robots, sim-to-real transfer.
Table of contents
- Key takeaways
- What the lab is and who funds it
- π0.5: the claim and the paper
- The ablation table, read carefully
- What openpi actually gives you
- Demonstrated versus claimed
- Partners and the commercial edge
- What to watch
- Frequently asked questions
- Changelog
What the lab is and who funds it
Physical Intelligence publishes at www.pi.website (the older physicalintelligence.company now 308-redirects there). Its stated thesis, from the openpi post, is that "general purpose models that can control any robot to perform any task will be the future of robotic control."
The homepage lists investors: Bond, Jeff Bezos, Khosla Ventures, Lux Capital, OpenAI, Redpoint Ventures, Sequoia Capital, CapitalG, and Thrive Capital. No amounts and no valuation appear on the site, so anything you have read about round sizes did not come from the company's own pages.
The π0.5 author list is unusually legible for a startup: 35 named authors, from Kevin Black to Ury Zhilinsky, including Danny Driess, Chelsea Finn, Karol Hausman, Brian Ichter, Sergey Levine, Karl Pertsch, and Quan Vuong. This is an academically-staffed lab that publishes on arXiv, which is why its evidence base is checkable at all.
π0.5: the claim and the paper
The paper is "π0.5: a Vision-Language-Action Model with Open-World Generalization," arXiv 2504.16054, submitted 22 April 2025, still at v1. The headline claim, verbatim from the abstract: "we demonstrate for the first time that an end-to-end learning-enabled robotic system can perform long-horizon and dexterous manipulation skills, such as cleaning a kitchen or bedroom, in entirely new homes."
The method is co-training on heterogeneous sources: data from multiple robots, high-level semantic prediction, web data, and "hybrid multi-modal examples that combine image observations, language commands, object detections, semantic subtask prediction, and low-level actions." In plain terms, the model predicts what subtask to do next in language and then predicts the motor actions for it, trained jointly.
Worth reading alongside the claim is the lab's own framing of the limits, stated on the π0.5 blog: "its goal is not to accomplish new skills or exhibit high dexterity, but to generalize to new settings." They also write that it "does not always succeed on the first try" and "often makes mistakes both in terms of its high-level semantic deductions and motor commands." That is candid, and it should recalibrate anyone who watched the videos and inferred a finished product.
The ablation table, read carefully
This is the part with real information density. Success rate is a per-subtask average (for example, the fraction of objects placed correctly); follow rate is the fraction of cases where behaviour accorded with the prompt.
| Condition | π0.5 (full mix) | no web data | no cross-embodiment | no multi-environment |
|---|---|---|---|---|
| In-distribution follow rate | 86% | 86% | 74% | 66% |
| In-distribution success rate | 83% | 82% | 67% | 57% |
| OOD follow rate | 94% | 80% | 67% | 33% |
| OOD success rate | 94% | 74% | 49% | 31% |
Two readings hold up. First, web data barely matters in-distribution (86/83 versus 86/82) and matters a great deal out of distribution (94/94 versus 80/74). Second, multi-environment data is the load-bearing ingredient: strip it and OOD success falls to 31%.
One caution on the table. π0.5's OOD scores exceed its in-distribution scores, which looks paradoxical. The two rows are different task sets, full home cleaning versus moving a prompted object into a drawer, so the OOD condition is a separate test and does not restate the same one at higher difficulty. Do not read the numbers as "it generalizes better than it interpolates."
The scaling study is the other durable finding: performance rises with the number of training environments, and "after only about 100 training environments, it actually approaches the performance of the baseline model that was trained on test environment directly." A fifth ablation combining "no ME or CE" is charted without published numbers; that condition retains "only the mobile manipulation data collected with the same robots that we use in our experiments (about 400 hours)."
What openpi actually gives you
openpi, announced 4 February 2025, is the release that made this lab a reference point. The repository carries an Apache 2.0 LICENSE file. Neither blog post states licence terms for the weights specifically, so the Apache 2.0 grant is verified for the code and not for the checkpoints. That gap is worth noting if you are doing procurement diligence.
The original release covered π0 base weights and code, fine-tuned ALOHA and DROID checkpoints, inference examples, and fine-tuning code. π0.5 weights came later: the README records "[Sept 2025] We released pi05, an upgraded version of pi0 with better open-world generalization," with a base checkpoint at gs://openpi-assets/checkpoints/pi05_base. The same README carries the caveat that matters most: "in this repository, we currently only support the flow matching head for both π0.5 training and inference." The open artifact is a subset of the paper system.
Practical numbers from the repo. Base checkpoints are "pre-trained on 10k+ hours of robot data." Inference needs more than 8 GB of GPU memory (an RTX 4090 clears it), LoRA fine-tuning more than 22.5 GB, full fine-tuning more than 70 GB (A100 80GB or H100). The repo has been tested on Ubuntu 22.04 only. On distributed training, the README's caveat is scoped: "the current training script does not yet support multi-node training," which applies to the JAX path; the PyTorch trainer documents a multi-node torchrun invocation. The blog's fine-tuning guidance is that "between 1 and 20 hours of data was sufficient to fine-tune to a variety of tasks," qualified with "your mileage may vary."
The single published simulation benchmark is LIBERO, π0.5 at 30k fine-tuning steps: Spatial 98.8, Object 98.2, Goal 98.0, Libero-10 92.4, average 96.85. The repo calls this state of the art. That is the lab's own characterization. Still, the checkpoint is public and the benchmark is standard, so this number is the one you can check yourself.
The lab's own disclaimer sets expectations: "This is an experiment... π0 may or may not work for you, but you are welcome to try it and see!"
Demonstrated versus claimed
Measured and reproducible. The LIBERO average of 96.85 against a published checkpoint. The π0.5 ablation table and environment-scaling curve, self-run but documented with method in a public paper.
Self-reported or partner-reported, not reproducible. Every real-home cleaning result. The Weave and Ultra intervention and grasp figures, written by those partners. π*0.6 and π0.7 success, throughput and deployment claims. "5x faster" training for the FAST tokenizer. The "about 4-5x higher" inference cost for π0-FAST versus flow matching, which the blog qualifies with "in our experience."
Positioning. "Control any robot to do any task," "bringing general-purpose AI into the physical world," "Cambrian explosion of robotics applications."
The absences are as informative as the presences. There is no total training-data scale for π0.5, only "about 400 hours" for one ablation subset and "10k+ hours" for openpi base checkpoints. No count of homes, no fleet size, no episode counts. No absolute real-world success rate for the headline home-cleaning claim: the blog shows videos and an ablation chart, with no per-task success table, no time-to-completion, no intervention rate. No inference latency or control-frequency figures for π0.5. Nothing on safety, force limits, e-stop behaviour, or what a failure looks like in an occupied home. And no commercial terms at all: no pricing, no availability, no SLA, no product name for a shippable system. The route in is "contact us."
Partners and the commercial edge
The π0.5 blog names no partners or customers. "The Physical Intelligence Layer" post, dated 24 February 2026, names two: Weave, which "builds robots for the home" and whose earlier commercial versions are "deployed in businesses across the San Francisco Bay Area" doing laundry folding, and Ultra, which builds industrial robots for warehouse work and describes "a fleet of revenue-generating robots deployed across the US, now scaling toward hundreds of deployments."
Read the framing on that page carefully. The page states that each partner section was written by our collaborators at Weave and at Ultra, so the numbers are partner-reported, and every one of them is a π0.6 result. π0.6 is a model with no dedicated post on the homepage index and no released weights.
Weave reports that π0.6 significantly increases autonomy compared to π0.5, and that adding Weave's data into pre-training "reduces missed grasps by 42% and interventions by 50%" per full laundry load. Ultra reports that π0.6 increases success rate versus π0.5, and that adding Ultra's pre-training data further increases order-packing throughput. The throughput chart carries no stated value. Its y-axis is unlabelled and the page publishes no measured figure, so there is no throughput number to quote. Treat all of this as a directional signal that the co-training recipe transfers to paying deployments.
The later models follow the same pattern. π*0.6 (17 November 2025) introduces an RL method called Recap. The lab reports success rates over 90% on the studied tasks, and says training with Recap on autonomous experience "more than doubles the throughput on some of the hardest tasks," naming espresso making as a case where "both throughput and success rate more than double from including on-robot experience." The lab also describes deployments in which π*0.6 made espresso drinks from 5:30am to 11:30pm, folded 50 novel laundry items in a new home, and assembled and labelled 59 boxes in a real factory. Those are demonstration claims, filmed by the lab, with no independent check. π0.7 (16 April 2026) is described by the lab as a steerable generalist prompted with language, metadata, control modalities, and visual subgoals; the lab reports running it on a bimanual UR5e system with Robotiq parallel-jaw grippers, and comparing it zero-shot on cross-embodiment laundry folding against teleoperators with a mean of 375 hours of experience. Both models have write-ups. Neither indicates a weights or code release.
What to watch
Whether openpi ever catches up to the flagship. A π0.7, π0.6 or π*0.6 checkpoint release would change the open-weights landscape for manipulation; continued silence tells you the lab has moved to a closed-frontier posture with an older open anchor.
Whether the flow-matching-only limitation on π0.5 in openpi is lifted, which determines how faithfully outside groups can reproduce the paper.
Whether any independent evaluation of a π model appears. Right now every real-world figure traces back to the lab or a partner.
Whether the environment-scaling result replicates elsewhere. "About 100 training environments approaches a directly-trained baseline" is the most portable claim in the whole body of work, and the one most worth an outside test.
Whether commercial terms surface. A product name, pricing, or an SLA would mark the shift from research lab to vendor.
Frequently asked questions
Is π0.5 still Physical Intelligence's best model? No. π0.7, dated 16 April 2026, is the current flagship on the site, and π0.6 and π*0.6 came before it. π0.5 remains the most recent model with released weights.
Can I download and run π0.5? Yes. The openpi repository's code is Apache 2.0 and the π0.5 base checkpoint was added in September 2025. Inference needs a GPU with more than 8 GB of memory; full fine-tuning needs more than 70 GB.
Is the open π0.5 the same as the paper's π0.5? Not entirely. The repo README states it currently supports only the flow matching head for π0.5 training and inference.
How much data was π0.5 trained on? The company does not publish a total. The blog gives "about 400 hours" for one ablation subset, and openpi's README describes base checkpoints as pre-trained on "10k+ hours of robot data."
Has anyone outside the lab verified the home-cleaning results? No independent evaluation of any π model appears on the pages reviewed here. The home results are presented as videos plus an ablation chart, without a per-task success table.
Why do the out-of-distribution scores look better than the in-distribution ones? Because they measure different task sets. The OOD condition is moving a prompted object into a drawer; the in-distribution condition is full cleaning tasks. Compare across ablation columns within a row.
Can I buy a Physical Intelligence system? Nothing on the site indicates a purchasable product: no pricing, availability, SLA, or product name. Named commercial relationships are with Weave and Ultra.
