Physical Intelligence, a hot robotics startup, says its new robot brain can figure out tasks it was never taught

The claim that π0.7 can perform untrained tasks is doing a lot of work here, but Physical Intelligence hasn't specified what evaluation methodology distinguishes genuine zero-shot generalization from interpolation across a very large training distribution. That distinction is the entire ballgame.

This lands in the middle of a crowded week for robot intelligence claims. Google DeepMind's Gemini Robotics-ER 1.6, covered here on April 13, made structurally similar promises about improved spatial reasoning and real-world manipulation, also without independent benchmarks at launch. MIT Technology Review's piece published the very next day traces exactly this pattern: robotics announcements consistently overpromise generality while delivering narrow, task-specific performance. Physical Intelligence is a better-funded and more credible player than most, but the announcement follows the same template those pieces describe. The honest read is that 'general-purpose' remains a directional aspiration, not a demonstrated property.

Watch whether Physical Intelligence publishes a technical report with held-out task categories and failure rates within the next 60 days. If the capability claim rests only on a demo reel and a blog post, treat it accordingly.