Agibot releases Physical Reasoning dataset for robot decision training

The dataset is organized around nine reasoning abilities—object selection, context‑based choice, feasibility assessment, prerequisite reasoning, goal‑directed search, state and quantity tracking, long‑horizon planning, memory, and subtask ordering—and is split into Easy, Medium and Hard difficulty levels. Annotations include dense prompts that map language instructions to bounding boxes, trajectories, and gripper states, as well as success and adversarial segments. AGIBOT says the data already supports its upcoming agent framework and foundation model, though performance results are not disclosed. The repository follows LeRobot conventions, provides head and wrist camera video streams, and is released under a Creative Commons Attribution‑NonCommercial‑ShareAlike 4.0 license.