AGIBOT WORLD 2026 Theme 3 opens 11,430 real-world robot trajectories for reinforcement learning

A successful trajectory shows a policy one route through a task. A failed trajectory reveals where that policy stops being reliable. That distinction matters for physical systems because a robot can reach the same failure state through bad perception, poor contact, an incorrect motion sequence or an unexpected change in the environment.

Theme 3 keeps those unsuccessful executions in the dataset. Researchers can inspect where a rollout diverged, how far the task progressed and what happened before a human intervention. That makes the data useful for failure recognition, recovery policies, reward modeling and value estimation rather than imitation alone.

AGIBOT divides the release into three complementary trajectory types. Expert demonstrations capture reference executions from human operators. Policy rollouts record autonomous model execution and preserve both successes and failures. Human-in-the-loop trajectories retain the robot's behavior before takeover, the intervention point and the corrective actions that follow.

That structure creates a continuous record from how a task should be performed to how a learned policy actually behaves on hardware. It also preserves the recovery path after something goes wrong, which is often missing from datasets built only from successful demonstrations.

The first batch spans 14 industrial and household tasks. AGIBOT names examples including Ethernet-cable insertion and unlocking a door with a key. These are useful reinforcement-learning cases because they require more than reaching a visual target. The robot has to manage contact, sequence actions and respond to execution feedback.

The release is therefore relevant to manipulation research where small errors accumulate. A cable can be aligned but not fully inserted. A key can reach the lock but fail to rotate correctly. Those partial states are important because they sit between a clean success label and a total failure label.

AGIBOT reports 98,159 annotated subtask intervals, 26,493 disturbance segments, 5,795 error-state segments and 10,684 human-intervention segments in the first batch. The annotations cover task progress, completion, errors, disturbances and interventions.

For researchers, that opens several directions. A model can be trained to estimate task progress, detect success, recognize errors, warn about risk or decide when human assistance is needed. The intervention segments can also support learning from corrective behavior rather than restarting every failed attempt from scratch.

Real-robot reinforcement learning is expensive because every experiment consumes hardware time and exposes the robot to wear, collisions and recovery work. Shared trajectories cannot replace new physical experiments, but they can reduce duplicated collection and give research teams common data for comparing methods.

The broader shift is from learning only what successful behavior looks like toward modeling what happens during execution. For humanoids and other general-purpose robots, reliable work will depend on recognizing when an action is going wrong and recovering before a small error ends the full task. Theme 3 gives researchers data specifically around that boundary.