Physical Intelligence, a startup founded by former members of Google Brain and DeepMind, is applying large language models to robotics to enable machines to understand and execute complex tasks with minimal human intervention.
The company leverages the same neural network architectures that power systems like GPT and Claude, but trains them on robot sensor data and motor control sequences rather than text. This approach allows robots to build generalizable understanding of physical interactions, physics principles, and task hierarchies from diverse training examples.
The startup's core innovation involves using language models as a foundation for what researchers call "embodied AI." Rather than programming specific behaviors for each task, robots learn from demonstrations and textual instructions to infer underlying principles. A robot trained this way can adapt to novel situations without explicit retraining.
Physical Intelligence's work addresses a longstanding robotics bottleneck. Traditional approaches required extensive hand-coding for each task. Deep learning methods improved flexibility but demanded massive labeled datasets specific to each robot and environment. By borrowing the transfer-learning capabilities of large language models, the startup aims to create robots that generalize across different hardware and task domains.
The company has demonstrated its approach on manipulation tasks involving object handling, assembly, and tool use. Early results show robots can understand context-dependent instructions, learn from corrective feedback, and apply learned knowledge to related but previously unseen tasks.
Challenges remain substantial. Robots still struggle with tasks involving long-term planning, physical reasoning about unseen objects, and real-world variability in lighting and surfaces. The computational demands of running large models on robot hardware also remain prohibitive for mobile systems.
Physical Intelligence operates within a rapidly expanding field. Competitors including Tesla's AI division and established robotics firms are pursuing similar paths. The startup's advantage lies in its team's track record in machine learning and its willingness to commit significant computational resources to the problem.
The approach represents a philosophical shift in robotics