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

2026-04-16 · Source: AI News & Artificial Intelligence | TechCrunch · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Advanced, medium

Summary

Physical Intelligence, a San Francisco-based robotics startup, has published new research on its latest model, π0.7, demonstrating compositional generalization. This model can direct robots to perform tasks not explicitly trained for, a significant step towards a general-purpose robot brain. Unlike traditional rote memorization approaches, π0.7 combines learned skills from different contexts to solve novel problems. A key demonstration involved an air fryer the model had minimal prior exposure to, yet it successfully attempted to cook a sweet potato with verbal instructions. The model also matched the performance of previous specialist models across tasks like making coffee and folding laundry. Researchers acknowledge limitations, including the inability to execute complex multi-step tasks autonomously from a single high-level command and the current lack of standardized robotics benchmarks. Physical Intelligence has raised over $1 billion and is reportedly seeking a new round that could value it at $11 billion.

Key takeaway

For entrepreneurs and investors evaluating robotics ventures, the π0.7 model's compositional generalization capability signals a potential inflection point in robotic AI, akin to large language models. Your focus should shift towards systems demonstrating true generalization over highly choreographed demos, as this indicates more scalable and useful real-world applications. Consider the long-term value of models that can adapt to new tasks with minimal training or human coaching, rather than those requiring extensive, task-specific data collection.

Key insights

The π0.7 model demonstrates compositional generalization in robotics, enabling robots to perform novel tasks by combining learned skills.

Principles

• Capabilities scale super-linearly with data once compositional generalization is achieved.
• Verbal coaching can improve robot performance in new environments without retraining.

Method

The π0.7 model synthesizes fragments from limited task-specific data and broader web-based pretraining to develop functional understanding for unfamiliar appliances.

In practice

• Refine prompt engineering for robot tasks to significantly increase success rates.
• Break down complex robot tasks into step-by-step verbal instructions for execution.

Topics

• Physical Intelligence
• π0.7 Model
• Compositional Generalization
• Robotic AI
• Robot Training

Best for: Investor, Entrepreneur, AI Scientist, Robotics Engineer, Research Scientist

Related on AIssential

• Physical Intelligence shows robot model with LLM-like generalization, flaws included — π0.7 is a robot foundation model exhibiting compositional generalization by recombining learned skills.
• Why Robots Still Struggle With Simple Tasks (And What Might Finally Change That) | Karol Hausman, Co-Founder & CEO of Physical Intelligence — Combining internet-scale LLM knowledge with robot motion data enables general physical intelligence for complex tasks.
• A robot operating system framework for using large language models in embodied AI — Connecting LLM agents to ROS enables versatile, robust, and scalable embodied AI through natural language.
• ATOM-Bench: A Real-World Benchmark for Atomic Skills and Compositional Generalization in Manipulation Policies — Current robotic manipulation policies struggle with fine motor skills and compositional generalization, even when atomic skills are strong.
• Announcing Raia Hadsell (Google DeepMind) at RAAIS 2026 — The article highlights Raia Hadsell's consistent focus on lifelong learning and embodied AI, bridging foundational research with frontier systems.
• Factored Diffusion Policies:Compositionally Generalized Robot Control with a Single Score Network — Factored Diffusion Policies use a single diffusion network with additive score decomposition for compositionally generalized robot control.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI News & Artificial Intelligence | TechCrunch.