AGIBOT releases 11,430 real-world robot trajectories

A robot's failed attempt is part of the record in AGIBOT's open-source release. The Chinese robotics company has published 11,430 real-world robot trajectories across 14 industrial and household tasks, capturing not only successful actions but also breakdowns and the moments when human operators step in.

The dataset targets reinforcement learning, in which robots improve from experience, and embodied AI, where learning is tied to interaction with the physical world. That makes it different from collections built mainly around successful expert demonstrations: AGIBOT's WORLD 2026 Theme 3 also records what happens when a learned policy struggles in an actual environment.

The records fall into three groups. Expert demonstrations show human operators completing tasks. Autonomous policy rollouts show robots attempting tasks independently, with 1,024 successful and 1,369 failed attempts included. Human-in-the-loop trajectories capture the robot's behavior before an operator takes control, followed by the corrective actions used to recover or complete the task.

The useful detail is in the annotation. AGIBOT says the dataset marks task progress, completion status, errors, environmental disturbances and human interventions. Researchers can therefore study not just whether a robot succeeded, but where its behavior failed and how guidance helped it recover—signals that clean demonstrations alone do not provide.

Concretely, the open-source release gives researchers data to study failures, interventions and recovery rather than successful demonstrations alone. That could help move research toward robots that learn more continuously and operate more reliably in everyday settings. The company says additional datasets, benchmarks and research resources are still planned.