Physical Intelligence: π0.7, a Robot with Generalization Ambitions

A Robotic Model Inspired by Language Models

The American start-up Physical Intelligence recently unveiled π0.7, a robotic model that draws inspiration from large language models to recombine acquired skills. This approach, termed "compositional generalization," allows the robot to combine fragments of skills learned during training, similar to how language models recombine text.

The π0.7 model is based on Google's Gemma3, which has four billion parameters, and is paired with a smaller action expert of 860 million parameters to generate the robot's actual movements. Physical Intelligence emphasizes that the key lies in the training method rather than the architecture itself.

Contextual Training and Flexibility

Unlike previous models that received brief instructions, π0.7 is fed a multitude of contextual information. This information includes natural language sub-task instructions, metadata on the quality and speed of demonstrations, as well as real-time generated images of sub-goals. This approach allows for the utilization of data of varying quality by simply tagging failed or slow attempts with appropriate metadata.

Previous robotic models typically received a short task description during training, such as "fold the t-shirt." In addition to this, π0.7 receives a range of contextual information: natural language sub-task instructions, metadata on the quality and speed of the demonstration, control mode labels, and even images of sub-goals showing what the result of an intermediate step should look like. These sub-goal images are generated in real-time by a second lightweight world model.

Performance and Challenges of Generalization

Physical Intelligence reports that a single π0.7 model matches the performance of the specialized π*0.6 models previously fine-tuned through reinforcement learning on tasks such as folding laundry, making espresso, and constructing boxes. Inter-embodiment transfer also works: a bimanual industrial manipulator UR5e folded t-shirts with an 80% success rate, even though no folding data had been collected for this robot. According to PI, this corresponds to the zero-shot performance of experienced human operators attempting the task on this robot for the first time.

New tasks can also be taught through linguistic coaching. A human guides the robot through the activity step by step by giving individual instructions. These coaching episodes can then be used to train a high-level policy that executes the task autonomously, without the need to collect conventional teleoperation data.

The Air Fryer and the Question of Compositional Generalization

As a primary example of compositional capability, PI cites loading a sweet potato into an air fryer. Without guidance, the model fails, but with step-by-step coaching, it succeeds. In the technical report, the team notes that they found only two episodes in the training data where a robot closes an air fryer, plus data from the open-source DROID dataset involving a Franka robotic arm.

However, a closer examination of the demonstration video reveals that the Franka arm from the DROID dataset opens an air fryer drawer and places a bottle inside. Structurally, this is very close to the sweet potato task that π0.7 is supposed to solve by recombining known skills. PI describes these episodes as "quite different" from what the mobile robot does in the experiment and interprets the result as evidence that the model composes new skills, just as language models recombine fragments of text from the web.

This raises a familiar debate from the world of language models in robotics: whether a model is genuinely solving a new task through generalization or essentially recalling very similar training data. With language models, this has been discussed for years under the term "data contamination," when evaluation tasks appear identically or in a very similar form in the training material.

PI itself admits in the report that given the enormous size and diversity of the dataset, it is difficult to determine with certainty which tasks are genuinely new. However, the team argues that this recombination of known building blocks is the essence of "compositional generalization." In practice, they claim there is no significant difference between a skill derived from generalization or transferred from similar situations (remixed, as they call it).

Language Model Phenomena Reach Robotics

π0.7 suggests that foundational robotic models are reaching a scale at which effects similar to those of large language models become visible: the nature of the prompt takes on considerable importance, performance heavily depends on the context provided, and distinguishing between "authentic generalization," remixing, and retrieving similar examples becomes the central evaluation issue.

Additional ablations in the report also show how important metadata is for scalability. Without quality annotations, the model deteriorates when more lower-quality data is added. With metadata, it continues to benefit from additional data even if the average quality decreases.

The report does not address the topic of reasoning models. PI only hints at the end that steerable models like π0.7 could in the future solve more complex tasks by "thinking" about possible approaches in advance. The current model does not yet take this step on its own.