AgiBot open-sources a decision-making robot dataset, starting with 150 hours of real-robot data

AgiBot on Oct. 22 open-sourced Physical Reasoning, the fourth installment of its AGIBOT WORLD 2026 dataset series and the first of seven releases in its "AI WAVE" run of embodied AI products. The company says the design target has moved from "manipulation difficulty" to "decision difficulty": the same task is recorded under different starting states, so restocking a hotel minibar to three bottles means telling empty from full and counting before deciding what to add, and a key in a box can be grabbed directly or only after the box is moved. Errors are labeled alongside successes.

The release doubles as a preview of AgiBot's own stack. AgiBot says the data will train its upcoming agent framework and a new foundation model; its schedule lists Genie Master, billed as connecting agents to the physical world, for Oct. 27 and the foundation model for Oct. 29. The first batch is 150 hours, a small amount for teaching judgment. The decision labels do come with real-robot action trajectories and box, point and line annotations, so the set can also train ordinary action policies.

AgiBot's own test comes on Nov. 2-3, when six G2 robots will work four shifts in a real convenience store, livestreamed for 24 hours. When the shelf count does not match the plan, will they count before they restock?