AGIBOT Releases Physical Reasoning Dataset to Connect Robot Decisions With Actions

AGIBOT has announced Physical Reasoning, the fourth release of its AGIBOT WORLD 2026 dataset, designed to help robots decide what to do as the situation around them changes. The October 22 announcement shifts the emphasis from difficult movements to decisions: which object to choose, what needs to happen first, and whether a task has actually been completed.

Key Takeaways

Show all

• Release 4 records different action paths for the same goal under changing conditions.
• It combines high-level questions and answers with robot action trajectories and detailed annotations.
• Tasks cover nine reasoning abilities, including search, quantity judgments, memory and subtask ordering.
• AGIBOT says it is already using the data for an upcoming agent framework and next-generation foundation model.

The same goal can require different actions

In a press release supplied to Humanoids Daily, AGIBOT illustrates the approach with a hotel minibar. To leave three full bottles, a robot must identify empty bottles, count the full ones and decide what to remove or add. Repeating a fixed sequence would not work across different starting conditions.

Another example involves searching drawers for toiletries. The next step depends on which drawers have already been checked, even when the current camera view looks the same. Retrieving a key may require moving an obstructing box first, while preparing items for an outing can involve checking a weather forecast before choosing sunscreen or an umbrella.

The AGIBOT WORLD project organizes this release around variations in object quantity, obstructions, available space and task progress. Its nine reasoning categories span selection, manipulation feasibility, action preconditions, search, state tracking, planning, memory and ordering subtasks. Tasks are divided into three difficulty levels, from individual judgments to combinations of goals and constraints.

Connecting decisions to robot movements

The annotations link an overall instruction to subtasks and individual actions. Detailed action descriptions specify the target object, where to manipulate it, how the two arms divide the work, gripper state and the expected result. Spatial labels connect those descriptions to locations and motion paths in the images.

AGIBOT says the release includes successful and erroneous examples, with annotations recording whether the correct object was selected or placed and whether quantity requirements were met. That gives researchers material for studying how a system detects a mistake and adjusts its subsequent actions.

The company says the dataset has already supported its forthcoming agent framework and training of its next-generation foundation model. The press release does not report a Release 4 episode count or quantitative improvements in robot task success.

Available for research, with license conditions

AGIBOT describes the release as open-source and links to the AGIBOT WORLD 2026 repository on Hugging Face. The repository currently lists a CC BY-NC-SA 4.0 license, which includes a non-commercial restriction.

The announcement provides a research resource for connecting decisions with physical execution. Its practical test will be whether models trained on these varied situations can choose the right next action in settings beyond the collected examples.