Physical Intelligence announced π0.7, a general-purpose robotics foundation model that can perform dexterous tasks without fine-tuning and recombine skills to solve novel problems. The model accepts language instructions, visual subgoals, and metadata annotations that describe not just what to do but how to do it—speed, quality, control modality.

The key technical advance is the prompting framework. By conditioning on diverse prompt types during training, π0.7 can integrate data from multiple robot platforms, human video, and autonomous policy rollouts—sources that would otherwise conflict due to differing strategies and proficiency levels. At inference, the model accepts standard language commands plus optional visual subgoals generated by a lightweight world model.

Physical Intelligence demonstrates π0.7 operating kitchen appliances it was never explicitly trained on, including an air fryer task performed via step-by-step language coaching. The company also shows cross-embodiment transfer: a bimanual UR5e system folding laundry despite no laundry-folding data collected on that platform. Training data included episodes from the open-source DROID dataset and proprietary home environment collections.

The announcement positions π0.7 as a step toward LLM-style compositional reasoning in physical systems. If the generalization holds across diverse real-world deployments, it represents a shift from task-specific specialist models to general-purpose robotics policies. Independent validation and deployment data will clarify whether the lab results translate to production environments.