Zhiyuan has recently taken a significant step by officially open-sourcing the third-phase dataset of AGIBOT WORLD 2026. This dataset is specifically tailored for real-machine reinforcement learning scenarios, boasting an impressive collection of 11,430 trajectories. These trajectories encompass 14 quintessential interaction tasks across two pivotal domains: industrial manufacturing and household services.
The dataset is meticulously curated from the full spectrum of robot operations, with each data point enriched by multi-dimensional, fine-grained annotations. In comparison to its predecessors, the first and second phases, this latest dataset significantly elevates the complexity of interaction scenarios within real-world settings.
It serves as a robust solution to the longstanding challenges of inadequate generalization and feeble robustness that have plagued the deployment of embodied intelligence research in practical applications. By doing so, it provides unwavering support and a solid foundation for advancing research in the field of embodied intelligence.