AGIBOT open-sources Physical Reasoning dataset as AGIBOT WORLD 2026 Release 4
SUMMARY
AGIBOT announced on 22 October 2026 that it has open-sourced the "Physical Reasoning" dataset, the fourth release of AGIBOT WORLD 2026. The company says this is the first release in the initiative to move the task design away from manipulation difficulty and towards the complexity of decisions, so that robots can choose and carry out actions as conditions change.
The dataset records how the same goal is handled under different states and constraints, with more than 10 categories of logical reasoning scenarios. AGIBOT names four factors that vary: object quantity, occlusion, spatial occupancy and task progress. Examples in the release include restocking a hotel-room minibar and retrieving a key that may sit behind a box. It provides high-level question-and-answer data together with low-level action trajectories, which AGIBOT says allows high-level decision models and low-level action policies to be trained together.
The dataset addresses nine reasoning abilities, among them goal-directed search, long-horizon planning, memory and state maintenance, and subtask ordering. Tasks come in three difficulty levels: Easy, Medium and Hard. Execution outcomes are annotated, and successful and erroneous samples carry equal weight, which AGIBOT presents as material for studying how systems adjust after deviations.
The annotations include a high-level instruction, a task frame, subtask segments and action steps. At action level, a Dense Prompt describes the target object, manipulation region, bimanual role assignment, motion process, gripper state and expected outcome. Bounding-box, point and trajectory-line annotations support visual-language grounding, and "success segments" and "adversarial segments" help assess task completion. AGIBOT says the dataset has already supported its upcoming agent framework and the training of its next-generation foundation model, with further progress to be announced.
Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.
WHY IT MATTERS
EDITORIALAGIBOT says the release pairs high-level question-and-answer data with low-level action trajectories, so decision models and action policies can be trained together. Successful and erroneous samples carry equal weight, giving researchers material to study how systems adjust after deviations during execution.