Video: Chinese firm releases 11,430 robot trajectories to advance research

The dataset captures successes, failures, and human corrections across the path from demonstration to deployment.AGIBOT/YouTube Chinese robotics player AGIBOT has open-sourced its WORLD 2026 dataset, which focuses on reinforcement learning for embodied AI. The dataset is designed to help robots learn from real-world interactions rather than relying solely on successful expert demonstrations. It captures the broader learning process, including how robots improve through failures and human intervention. Recently, AGIBOT topped the second World Humanoid Robot Games with 46 medals, including 18 golds, making its international competition debut. Real-world robot dataset AGIBOT’s WORLD 2026 Theme 3 dataset initiative provides researchers with real-world robot experience data aimed at advancing reinforcement learning and embodied AI. The open-source dataset contains 11,430 real-world trajectories collected across 14 industrial and household tasks. Unlike datasets focused primarily on successful demonstrations, the new release captures a wider range of robot experiences, including successful executions, failed attempts, autonomous deployments, and interventions by human operators. The dataset is organized around three types of trajectories. The first consists of expert demonstrations, which document reference task executions carried out by human operators in real-world environments. These demonstrations provide robots with examples of how tasks can be performed and serve as a foundation for learning. The second category covers autonomous policy rollouts. These records capture robots attempting tasks independently after deploying learned policies. Both successful and unsuccessful attempts are included, allowing researchers to examine not only what a robot can accomplish but also where and why its performance breaks down. The dataset includes 1,024 successful policy rollouts and 1,369 failed rollouts. The third category focuses on human-in-the-loop corrections. These trajectories document a robot’s behavior before an intervention, the moment a human operator takes control, and the corrective actions used to recover or complete the task. This provides researchers with data on how robots respond when autonomous behavior goes wrong and how human guidance can help overcome failures. AI learns failure AGIBOT has also included detailed annotations covering task progress, completion status, errors, environmental disturbances, and human interventions. Such information is intended to give researchers a more structured view of robot behavior and provide useful signals for reinforcement learning and other embodied AI applications. The release reflects a broader shift in robot learning from relying primarily on curated demonstrations toward learning from actual deployment experience. In real-world environments, robots can encounter unexpected obstacles, make mistakes, deviate from intended actions, or require assistance. Capturing these situations can help researchers study capability limits, failure modes, and recovery strategies that successful demonstrations alone may not reveal. Theme 3 forms part of the broader AGIBOT WORLD 2026 initiative, which is being developed as an open-source resource for embodied AI research. The company plans to expand the initiative with additional datasets, benchmarks, and research resources. According to the firm, by making real-world successes, failures, and corrective interactions available to researchers, it aims to provide a broader foundation for developing robots capable of learning continuously from their own experiences and operating more reliably in everyday environments. “We invite researchers worldwide to leverage AGIBOT WORLD 2026 to drive robotic intelligence from the lab into the real world, empowering every industry and tangibly boosting production and service efficiency,” reads the AGIBOT WORLD website. Get the latest in engineering, tech, space & science - delivered daily to your inbox.Jijo is an automotive and business journalist based in India. Armed with a BA in History (Honors) from St. Stephen's College, Delhi University, and a PG diploma in Journalism from the Indian Institute of Mass Communication, Delhi, he has worked for news agencies, national newspapers, and automotive magazines. In his spare time, he likes to go off-roading, engage in political discourse, travel, and teach languages.