The collaboration between Boston Dynamics and the RAI Institute aims to tackle key challenges in reinforcement learning for robotics, focusing on three main areas: sim-to-real transfer, whole-body loco-manipulation, and full-body contact strategies.
Despite advances in fast parallel simulators and optimization techniques, transferring simulation-trained policies to real robots remains a major hurdle. To bridge this gap, the teams will develop reinforcement learning policies that generate agile and adaptable behavior on physical hardware, enabling robust and practical locomotion.
Another focus is enhancing whole-body loco-manipulation, where robots must seamlessly combine locomotion with object interaction, such as opening doors or operating levers. By improving policy robustness in these scenarios, the teams aim to increase the practical utility of humanoid robots in real-world environments.