π0.7 Enables Robots to Remix Skills for New Tasks
Physical Intelligence's π0.7 model combines sparse training data into novel robot behaviors like air fryer use, succeeding with verbal coaching and scaling superlinearly like LLMs.
Compositional Generalization Unlocks Superlinear Scaling
π0.7 shifts robotics from rote memorization—training specialist models per task—to compositional generalization, where the model recombines skills across contexts for unseen problems. This mirrors LLM scaling: capabilities grow faster than data volume once generalization kicks in. Train on fragments like pushing an air fryer door (one episode) and inserting a bottle (one open-source clip), plus web pretraining, and it infers full appliance use. Researchers note data efficiency jumps, enabling deployment without per-task retraining.
Surprising Demos from Minimal Data
With zero-shot attempts, π0.7 handles novel objects like cooking a sweet potato in an untrained air fryer. Add step-by-step verbal coaching—like instructing a new hire—and success hits 95% (up from 5% via refined prompts). It matches prior specialist models on coffee-making, laundry folding, and box assembly. Even ad-hoc tests surprise creators: given random gears, it rotates them flawlessly. Balakrishna, knowing the dataset intimately, admits rare shocks, akin to GPT-2 inventing 'unicorns in the Andes' from thin air. Generalization prioritizes utility over flashy stunts like backflips.
Prompting Matters, Autonomy Lags
Failures often stem from poor instructions, not the model—half-hour prompt tweaks boost rates dramatically. It excels with walkthroughs ('open this, push that') but falters on single high-level commands like 'make toast.' No standard benchmarks exist, so validation relies on internal specialist baselines. Lacks full multi-step autonomy. Deployment timelines undisclosed, but progress outpaces expectations.
Startup Fuels Optimism
Physical Intelligence, 2-year-old SF firm, raised over $1B at $5.6B valuation, eyeing $11B round. Backed by Lachy Groom (early Figma/Notion investor), it draws institutional capital sans firm commercialization dates.