AGIBOT releases Physical Reasoning dataset to teach robots decisions

AGIBOT on October 22 announced Physical Reasoning, Release 4 of its AGIBOT WORLD 2026 dataset. The release links task goals, changing scene conditions, execution history and real robot actions, and includes both successful and erroneous examples.

• Release covers nine abilities, from object selection to long-horizon planning and memory
• Tasks are split into Easy, Medium and Hard levels by decision complexity
• Dense Prompt annotations specify target object, manipulation region, arm roles and gripper state
• Data was collected on the G2 platform under a CC BY-NC-SA 4.0 license