Physical Intelligence π 0.7 generalizes, steers robots

This looks less like a flashy demo and more like a real sign that robot foundation models are starting to move from imitation toward compositional control.

• –The interesting part is not just new tasks, but the promptable control surface: task language, strategy hints, metadata, and subgoals all become knobs for steering behavior.
• –If the results hold up outside curated demos, this could reduce how often teams need per-task finetuning or teleoperation-heavy data collection.
• –The caveat is obvious: robotics still lives or dies on latency, safety, and failure recovery, and the blog itself shows zero-shot attempts that only partly complete harder tasks.
• –For developers, the bigger implication is that robot policies may start to look more like multimodal systems engineering than classic control, with data curation and prompt design doing a lot of the work.