Physical Intelligence π0.7 Robot Model Shows LLM-Like Generalization
US startup Physical Intelligence launched π0.7, a new foundation model for robots. This system recombines skills from training data in a way that mirrors how language models piece together text. Researchers call these early indicators of compositional generalization in robotics.
Physical Intelligence focuses on creating versatile AI for physical tasks. The company builds generalist models that work across robot types and activities. Their prior releases, like π0.6, set the stage for this advance.
Model Architecture and Key Ingredients
π0.7 uses Google's open Gemma3 language model, which has four billion parameters. It pairs this with an 860-million-parameter action expert to produce robot movements. Physical Intelligence stresses that the architecture matters less than the training method.
Past robot models often get just a brief task description, like "fold the t-shirt." π0.7 receives extra details during training. These include natural language subtask instructions, episode metadata on demonstration quality and speed, control mode labels, and subgoal images for intermediate steps. A lightweight world model creates these subgoal images on the fly.
This setup allows training on data of mixed quality. Instead of throwing out failed or slow demos, teams tag them with metadata.
Strong Results from One Generalist Model
A single π0.7 instance equals the performance of earlier RL-fine-tuned π*0.6 specialist models. It handles laundry folding, espresso making, and box building equally well. The model also transfers across robot bodies. For instance, a bimanual UR5e industrial arm folded t-shirts at an 80 percent success rate. No folding data existed for this robot before. Physical Intelligence says this matches zero-shot results from skilled human teleoperators trying the task fresh.
New activities come through language coaching. A person guides the robot step by step with spoken instructions. These sessions train a high-level policy for independent operation. No traditional teleoperation data collection is needed.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
Air Fryer Test Highlights Composition Questions
Physical Intelligence points to loading a sweet potato into an air fryer as proof of compositional skill. Without coaching, the model fails. With step-by-step guidance, it succeeds. The technical report notes just two training episodes where a robot closes an air fryer. It also draws from the open-source DROID dataset, which features a Franka robot arm.
Demo footage shows the Franka arm opening an air fryer drawer and inserting a bottle. This setup closely resembles the sweet potato task. Physical Intelligence views these as distinct from the mobile robot's demo. They see the outcome as the model freshly combining skills, akin to language models reusing web text snippets.
This sparks a familiar robotics debate, borrowed from language AI. Does the model truly generalize to novel problems, or does it pull from near-identical training cases? Language model talks often frame this as data contamination, where test items match training material closely.
Challenges in Proving Novelty and LLM Parallels
Physical Intelligence admits the dataset's vast size and variety make it tough to confirm task novelty. Still, they argue recombining familiar elements defines compositional generalization. In real use, they claim it does not matter if a skill stems from pure generalization or adaptation from close examples, which they term remixing.
π0.7 points to robot foundation models hitting scales where language model traits emerge. Prompt details grow critical. Results hinge on supplied context. Key challenges involve separating true generalization, remixing, and example retrieval.
Report ablations confirm metadata's role in growth. Without quality tags, performance drops as low-quality data volume rises. With tags, benefits persist even as data quality averages down.
The report skips reasoning models. Physical Intelligence suggests steerable systems like π0.7 might one day tackle tough tasks by planning ahead. The current version does not do this independently.