Researchers release the general-purpose robot foundational model π0.7, achieving compositional generalization ability

ME News Report, April 17 (UTC+8), researchers recently released a new general-purpose robot foundational model called π0.7, which reportedly achieves significant breakthroughs in generalization ability. The model can perform a wide range of dexterous manipulation tasks, with performance comparable to specially fine-tuned expert models, and can understand new language instructions to complete tasks not seen in the training data, such as using new kitchen appliances or having an untrained robot fold clothes. π0.7 can execute all skills out of the box and combine them to solve new tasks, effectively generalizing across different robot platforms, scenarios, and tasks. The article suggests that the key to achieving generalization lies in using a broad and diverse dataset from different robots, humans, and autonomous strategies, and eliminating behavioral ambiguity by adding diverse contextual information in prompts (such as task text descriptions, visual sub-goal images, expected segment lengths, control mode labels, etc.), thereby integrating a wider range of data sources. (Source: InFoQ)