AGIBOT has released Physical Reasoning as open source, the fourth release in its AGIBOT WORLD 2026 initiative. Earlier releases concentrated on how hard a manipulation was to perform. This one turns to decision-making, so that robots can pick the right action and carry it out when real-world conditions shift.
AGIBOT describes AGIBOT WORLD 2026 as the first open-source dataset to span the full range of embodied artificial intelligence research. It is built around five research themes, each with its own data collection method and detailed annotation system for researchers in different subfields. Release 4 draws on complete real-world task requirements and records how the same goal is handled across different states and constraints. The aim is to move robots beyond isolated manipulations and toward finishing complex tasks on their own.
How a Robot Reasons Through an Everyday Request
Take the instruction “Help me pack what I need for going out tomorrow.” An agent first calls a tool to look up tomorrow’s weather. It then works out which items are needed, and a robot picks and places them. A sunny forecast means sunscreen goes in the bag, while expected rain means an umbrella does.
Release 4 ties together goal comprehension, task-state tracking and action execution. It covers more than 10 categories of logical reasoning scenarios and deliberately includes several real-world states and execution branches for each task. By pairing high-level question and answer data with low-level action trajectories, it supports training decision-making models and action policies together. The goal is for robots to “reason clearly” and “execute successfully.”
Project Page: https://agibot-world.com
Open Source Link: https://huggingface.co/datasets/agibot-world/AgiBotWorld202
Why Fixed Action Sequences Fall Short
Real tasks rarely follow a script. Restocking a hotel minibar to three full bottles means telling empty bottles from full ones, counting what is already there, and deciding how many to take out and how many to add. The right sequence depends on the starting state.
Searching for toiletries across several drawers is a different kind of problem. The camera view may stay the same, but the next move depends on which drawers the robot has already checked. Task progress, dependencies between steps and outside conditions all shape everyday jobs such as housekeeping, cooking and getting ready to leave the house.

Same Task, Different States
The dataset records decision-making and execution for one task under varying conditions, with a focus on four factors, namely object quantity, occlusion, spatial occupancy and task progress. Retrieving a key from behind a box shows the idea. If there is room to reach it, the robot grasps the key directly. If the box blocks the way, the robot must move the box first. The goal stays fixed while the action path changes.

Release 4 Targets Nine Reasoning Abilities
- Attribute-constrained selection
- Context-conditioned selection
- Manipulation feasibility assessment
- Action precondition reasoning
- Goal-directed search
- State and quantity reasoning
- Long-horizon planning
- Memory and state maintenance
- Subtask ordering reasoning

These skills often overlap. Minibar restocking calls for product discrimination, quantity judgment and state maintenance together, while finding a missing item involves search strategy, search history and spatial conditions.
Tasks are sorted into Easy, Medium and Hard levels. Easy tasks test single judgments, Medium tasks add quantity requirements and action preconditions, and Hard tasks combine multiple subgoals, long-horizon state tracking and several contextual constraints. Researchers can therefore isolate individual abilities or measure how they perform in combination.

Outcomes are annotated too, including whether objects were selected and placed correctly and whether the required quantity was reached. Successful and erroneous samples are weighted equally, which gives researchers material for studying how systems recover from mistakes during execution.
Annotations That Link Decisions to Actions
Each high-level instruction describes the full goal, while the task frame, subtask segments and action steps capture the structure of the task. Together they let researchers connect progress to the current subgoal and study task breakdown, stage recognition and next-subgoal prediction.

At the action level, a Dense Prompt gives a detailed description covering the target object, manipulation region, bimanual role assignment, motion process, gripper state and expected outcome, all aligned with real robot actions. Bounding-box, point and trajectory-line annotations tie language instructions to specific visual locations and motion paths. “Success segments” and “adversarial segments” help assess task completion, separate targets from distractors and check results.
What Researchers Can Do With It
The dataset supports joint or end-to-end training to improve situational understanding, process memory and task reasoning. It also enables decision and execution co-training. In that setup, an agent picks a subgoal based on the current situation and its history, the action policy performs the manipulation, and the task state is updated so later plans can adapt to the result.

By linking task decisions, process memory and real-robot execution, the dataset gives foundation models and embodied agents a way to turn an understanding of user goals into physical action. Over time, AGIBOT says this approach could help robots take part in full workflows in home services, hospitality and commercial settings, where tasks demand continuous judgment. Example scenarios include table resetting, retail restocking and guest-room cleaning.

Physical Reasoning is already supporting AGIBOT’s upcoming agent framework and the training of its next-generation foundation model. The company says further progress will be announced in due course.
Availability
Project page at agibot-world.com
Dataset on Hugging Face at huggingface.co/datasets/agibot-world/AgiBotWorld2026





