π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities

Authors: Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, Thomas Charbonnier, Vedant Choudhary, Foster Collins, Ken Conley, Grace Connors, James Darpinian, Karan Dhabalia, Maitrayee Dhaka, Jared DiCarlo, Danny Driess, Michael Equi, Adnan Esmail, Yunhao Fang, Chelsea Finn, Catherine Glossop, Thomas Godden, Ivan Goryachev, Lachlan Groom, Haroun Habeeb, Hunter Hancock, Karol Hausman, Gashon Hussein, Victor Hwang, Brian Ichter, Connor Jacobsen, Szymon Jakubczak, Rowan Jen, Tim Jones, Gregg Kammerer, Ben Katz, Liyiming Ke, Mairbek Khadikov, Chandra Kuchi, Marinda Lamb, Devin LeBlanc, Brendon LeCount, Sergey Levine, Xinyu Li, Adrian Li-Bell, Vladislav Lialin, Zhonglin Liang, Wallace Lim, Yao Lu, Enyu Luo, Vishnu Mano, Nandan Marwaha, Aikys Mongush, Liam Murphy, Suraj Nair, Tyler Patterson, Karl Pertsch, Allen Z. Ren, Gavin Schelske, Charvi Sharma, Baifeng Shi, Lucy Xiaoyang Shi, Laura Smith, Jost Tobias Springenberg, Kyle Stachowicz, Will Stoeckle, Jiaming Tang, Jimmy Tanner, Shalom Tekeste, Marcel Torne, Kyle Vedder, Quan Vuong, Anna Walling, Haohuan Wang, Jason Wang, XuDong Wang, Chris Whalen, Samuel Whitmore, Blake Williams, Charles Xu, Sukwon Yoo, Lili Yu, Wuming Zhang, Zhuoyang Zhang, Ury Zhilinsky · CoRL 2026 2026

Why it matters

Physical Intelligence released its earlier models, π0 and π0.5, as open code and weights in its openpi repository, and π0.7 is the newest model in that line. Its argument is about data rather than architecture: if every training episode is labelled with how it was performed (its speed, its quality, whether it contained a mistake, and a picture of the next intermediate goal), then failed runs, autonomous rollouts of earlier models, other robots and human video stop pulling the policy towards average behaviour and become usable training data. The authors report that one checkpoint prompted this way does what used to need a fine-tuned specialist for each task, follows unusual instructions in rooms it has never seen, can be talked through new appliance tasks step by step, and folds shirts on two industrial UR5e arms for which no folding data was ever collected. CoRL 2026 selected the paper as an oral. Readers cannot run it: π0.7 is not in openpi, and no code, weights or data have been released, so this page covers the method and the evidence, all of it reported by the authors.

Problem

Robot foundation models are trained mostly on good teleoperated demonstrations, with a short task description as the only prompt. That leaves out much of the data a lab actually produces: slower or sloppier demonstrations, failed episodes, autonomous runs of earlier policies, other robots and human video. Mixed in naively, data with different strategies and quality makes a policy average between them and perform worse. The same models also struggle to follow instructions unlike those in training, to carry a skill to a robot that never practised it, and to do even trained dexterous tasks well without being fine-tuned for each one. The paper asks whether a richer prompt can make all of that data usable and give one model these abilities without per-task fine-tuning.

Proposed solution

π0.7 keeps the flow-matching design of π0.6 with the MEM video memory: a Gemma 3 vision-language backbone of 4B parameters, a video encoder that compresses up to six past frames per camera, and an action expert of 860M parameters that outputs chunks of 50 actions, about 5B parameters in all. What changes is the prompt. Besides the task, it can hold the next subtask in words, subgoal images showing what each camera should see a few seconds later, episode metadata (episode length in bins of 500 steps, a quality score from 1 to 5 and a mistake flag) and the control mode, joint or end-effector. Each part is randomly dropped in training, so any subset works at test time. At run time the metadata asks for fast, top-quality, mistake-free behaviour; subtasks come from a high-level policy built on the same model or from a person coaching it; and subgoal images come from a world model fine-tuned from the 14B BAGEL image model, refreshed every 4 seconds or at each new subtask. Classifier-free guidance on the metadata pushes the actions further towards fast, clean execution.

Key insights

Labelling each episode with its speed, quality and mistakes turns failed and autonomous runs into useful training data; at test time the model is simply asked for fast, top-quality, mistake-free behaviour
Rollouts of π0.6* specialists trained with reinforcement learning can be distilled into the generalist this way: without the autonomous evaluation data, or without the metadata, results on the same tasks dropped, most of all in throughput
Subgoal images from a separate world model help most where language alone is weakest: instructions that go against habits in the data, and moving a task to a robot with a different body
A model that follows language closely can be coached through a new long task step by step, and those coaching sessions can then train a high-level policy that runs the task without a person and without new teleoperated data
Transfer to a very different robot meant changing strategy, not copying motions: on the taller UR5e arms the policy grasped cloth vertically instead of with the tilted grasp operators use on the source robot, and bagged a shirt with one arm instead of two

Main results

Folding a t-shirt on a bimanual UR5e system with no laundry data for that robot, π0.7 with generated subgoal images reached 85.6% task progress and an 80% success rate, against 90.9% and 80.6% for 10 expert teleoperators (about 375 hours of experience on average) on their first attempts, 30 trials in all
On the four tasks from the π0.6* release (two laundry tasks, espresso and box building), one π0.7 checkpoint matched the success rates of the per-task RL specialists and had higher throughput on diverse laundry and box building; on six further dexterous tasks it matched fine-tuned π0.6 specialists
Across 14 instruction-following scenarios of 3 to 6 steps in 4 unseen kitchens and 2 unseen bedrooms, π0.7 followed a clearly larger share of instructions than π0.5 and π0.6; the paper shows this as a chart without printed values
In the data-quality study, 8 models trained on the top 30%, top 50%, top 80% and all of the laundry data showed π0.7 with metadata improving as data grew, while the version without metadata got worse
The authors put success on seen tasks often above 90% and on unseen tasks or unseen task-robot pairs at 60 to 80%. Inference takes 38 ms in the minimal setup and up to 127 ms with memory and subgoal images on one NVIDIA H100; each subgoal image takes 1.25 seconds on 4 H100 GPUs

Limitations

No code, weights or data: as of 8 November 2026 openpi holds only the π0, π0-FAST and π0.5 checkpoints, Physical Intelligence's Hugging Face organisation only the FAST tokenizer and two datasets, and open questions in the openpi repository about releasing π0.6 and π0.7 have no answer from the team. The Pi API beta announced in October 2026 offers the company's latest models for hosted fine-tuning on application, without naming π0.7
Most results are bar charts without printed values, on tasks, robots and scoring rubrics of the authors' own; the only exact figures in the text are the shirt-folding comparison with teleoperators and the timings
Neither the size nor the make-up of the training data is given, nor the training compute, so the claims about data diversity cannot be checked or repeated
The authors say it is hard to tell which test tasks are truly unseen in so large a dataset, and unseen tasks succeed at 60 to 80% against often over 90% for seen ones; a long new task such as cooking a sweet potato in an air fryer did not work from a single prompt and needed step-by-step coaching
Running the whole system needs data-centre hardware: an NVIDIA H100 for the policy and the high-level policy, and 4 H100 GPUs for the 14B world model that draws the subgoal images

Field impact

The paper's claim is that prompt design, not a new architecture, lets a robot foundation model use mixed-quality data and begin to recombine skills. Its controlled experiments support the data part most directly: with metadata, adding more laundry data kept improving the model even as the data's average quality fell, while the same model without metadata got worse; removing the most task-diverse fifth of the data hurt unseen tasks more than removing a random fifth. The cross-embodiment result is the most striking: on UR5e arms with no laundry data, the model came close to expert teleoperators trying that robot for the first time. The authors note themselves that with this much data it is hard to say which tasks are truly unseen. Because neither the model nor the data is public, others can adopt the recipe, labelling episodes by quality and speed and conditioning on subgoal images, but cannot check the results; the paper mainly shows where Physical Intelligence has taken its models since the open π0 and π0.5.

Abstract

We present a new robotic foundation model, called π0.7, that can enable strong out-of-the-box performance in a wide range of scenarios. π0.7 can follow diverse language instructions in unseen environments, including multi-stage tasks with various kitchen appliances, provide zero-shot cross-embodiment generalization, for example enabling a robot to fold laundry without seeing the task before, and perform challenging tasks such as operating an espresso machine out of the box at a level of performance that matches much more specialized RL-finetuned models. The main idea behind π0.7 is to use diverse context conditioning during training. This conditioning information, contained in the prompt, makes it possible to steer the model precisely to perform many tasks with different strategies. It is conditioned not just on a language command that describes what it should do, but on additional multimodal information that also describes the manner or strategy in which it should do it, including metadata about task performance and subgoal images. This enables π0.7 to use very diverse data, including demonstrations, potentially suboptimal (autonomous) data including failures, and data from non-robot sources. Our experiments evaluate π0.7 across numerous tasks with multiple robot platforms, on tasks that require speed and dexterity, language following, and compositional task generalization.

Open Source Robotics