Can π0.7 Unlock Compositional Generalization and Cross‑Embodiment Transfer for VLA?

Two weeks after Generalist AI released Gen‑1, Physical Intelligence introduced the π0.7 model, pushing the visual‑language‑action (VLA) foundation forward.

π0.7 shows the first signs of compositional generalization: like a person who can combine chopping, heating, and stirring to make tomato‑egg stir‑fry without prior practice, the model can merge learned skills to solve unseen tasks.

In an air‑fryer experiment, the model had never seen the specific task “air‑fryer roast sweet potato.” By providing step‑by‑step language instructions (e.g., “close the basket,” “place food”), π0.7 understood and executed the full procedure, recombining concepts from different data fragments.

After a few language‑guided sessions, researchers fine‑tuned a high‑level policy that let π0.7 generate its own sub‑goals, completing the air‑fryer task autonomously, demonstrating the ability to stitch together scattered behavior snippets into a coherent action sequence.

To trace the knowledge source, the team identified two household video clips—one showing the basket being closed, another showing the basket placed on the left—and a segment from the open‑source DROID dataset with a Franka arm. Although these snippets differ from the robot’s execution, π0.7 recombines them rather than merely copying a single trajectory.

For cross‑embodiment transfer, π0.7 was tasked with folding clothes using an unseen dual‑arm UR5e system (two heavy UR5e arms with a Robotiq parallel gripper). Despite the robot’s large inertia and imprecise gripper, π0.7 achieved a zero‑shot success rate comparable to expert tele‑operators with 375 hours of experience.

The team previously built a task‑specific Recap algorithm that used reinforcement learning to improve speed and stability. Instead of training a separate Recap expert for each task, they distilled the Recap experience and policy metadata into π0.7. After distillation, π0.7 matched or exceeded the Recap experts on folding, coffee making, and box‑folding tasks.

Architecturally, π0.7 extends the π0.6 VLA backbone with a MEM memory system and adds multimodal context modulation. It incorporates a Gemma‑3 4B vision‑language model (0.4 B visual encoder) and a 0.8 B flow‑matching action expert, totaling roughly 5 B parameters.

The model relies on a diverse prompt framework that includes textual task descriptions, visual sub‑goal images, desired execution speed, metadata about action quality, control‑mode tags (joint vs. end‑effector), and visual sub‑goal pictures generated by a lightweight world model. This rich annotation lets π0.7 safely ingest low‑quality autonomous data by labeling it (e.g., “low quality”, “slow”).

Task and step‑by‑step language instructions

Metadata describing execution style (speed, quality)

Control‑mode tags

Visual sub‑goal images

The authors conclude that massive, diverse data combined with correct contextual prompts naturally give rise to surprising compositional abilities, turning many previously “hard” problems into tractable ones. Future work will focus on scaling data, improving evaluation metrics, and refining the role of lightweight world models.