Physical Intelligence's π0.5 cleans kitchens and bedrooms in homes it never saw in training, with 97.6% of its data from other sources

Physical Intelligence, the San Francisco robot-software start-up, on 22 April 2025 published π0.5, a vision-language-action (VLA) model that it says can clean kitchens and bedrooms in homes it has never seen. In a blog post and an accompanying preprint on arXiv, the company showed two-armed mobile robots putting dishes in a sink, making beds, wiping spills and clearing bedroom floors in three real homes that were excluded from training.

The figures in the paper are specific. The mobile robots learned from about 400 hours of household data collected in about 100 homes, yet the paper states that 97.6% of training examples in the first training phase did not come from mobile manipulators doing household work. The rest of the mix came from static robot arms in homes, laboratory robots including the public Open X-Embodiment collection, web data for captioning, question answering and object localisation, hand-labelled subtasks, and a small set of spoken coaching sessions. Single tasks took about 2 to 5 minutes, and full room clean-ups ran 10 to 15 minutes, according to the authors.

The architecture builds directly on π0, the company's first generalist policy published on 31 October 2024. π0.5 starts from a 2-billion-parameter vision-language model initialised from Google's PaliGemma weights and adds a 300-million-parameter "action expert". At each step the model first writes out, in text, the next subtask it should perform, such as picking up a pillow, and then generates motor commands for it. Those commands come out as 50-step action chunks covering one second of motion, produced by flow matching with 10 denoising steps. Training ran in two stages, 280,000 gradient steps on discrete tokens followed by 80,000 steps that add the continuous flow-matching head, the paper says.

The robots themselves are modest. The paper describes two mobile platforms, each with two 6-degree-of-freedom arms fitted with parallel-jaw grippers, a wheeled holonomic base, a torso lift and four cameras, two of them on the wrists. The state and action space has 18 or 19 dimensions depending on the platform. The company said its goal was not new skills or high dexterity but generalisation to new settings.

The most useful result for anyone building such systems is the company's ablation study. In a language-following test in held-out homes, the full π0.5 followed instructions 86% of the time and succeeded 83% of the time on familiar object types, and 94% on both measures for novel object types, according to the blog post. Removing web data barely moved the familiar-object score but cut success on novel objects to 74%. Removing laboratory cross-embodiment data cut novel-object success to 49%, and removing data from static robots in many homes cut it to 31%.

Scale of environments mattered too. The authors trained versions on mobile-robot data from 3, 12, 22, 53, 82 and 104 locations and found performance generally rose with more locations. The 104-location model performed about as well as a control model trained directly on data from the test homes, a result the blog post summarised as approaching that baseline after only about 100 training environments. The paper also reports that π0.5 significantly outperformed π0 and an enhanced π0 variant, even when π0 was trained for up to 300,000 steps, and that the full model beat a version in which a human supplied the high-level subtasks, while a zero-shot GPT-4 planner did worst.

Physical Intelligence is not alone. Figure AI published its Helix VLA on 20 February 2025, pairing a 7-billion-parameter model at 7 to 9 Hz with an 80-million-parameter controller at 200 Hz trained on about 500 hours of teleoperation. Google DeepMind followed on 12 March 2025 with Gemini Robotics, built on Gemini 2.0, and NVIDIA released its open GR00T N1 humanoid model on 18 March 2025. Where those efforts lean on a single body or a single owner's hardware, π0.5 is a test of whether data from many robots can be pooled into one model that works in someone else's house.

The company was candid about the limits. Its paper lists unfamiliar drawer handles, cabinets that are physically hard to open, an arm hiding the spill it was meant to wipe, and a high-level planner that sometimes opened and closed the same drawer repeatedly. The evaluation used three homes and averaged scores over 10 trials per task, which is a small sample, and all results are the company's own. Physical Intelligence later released π0.5 checkpoints in its Apache-2.0 openpi repository in September 2025, including a base model and versions fine-tuned on the LIBERO benchmark and the DROID dataset, which lets outside labs check the claims.