Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study enhances mobile manipulation behaviors in reinforcement learning by integrating 3D spatial information into a distance-aware world model, which provides more accurate reward signals for policy training. The authors demonstrate that this distance-based approach significantly outperforms traditional similarity-based rewards, achieving near 100% success rates in motion planning tasks, while similarity rewards only reach 60% success in fixed-target scenarios. Additionally, the use of reachability-based rewards for embodiment selection leads to over 85% rational arm activations, contrasting sharply with the 60% maximum of heuristic composite rewards.
Distance-aware rewards can boost mobile manipulation success rates to nearly 100%, leaving traditional methods in the dust.
Separate control of mobile manipulation offers a promising paradigm for managing its complexity by planning for these modalities independently. However, end-to-end learning of such behaviors necessitates effective mechanisms for both embodiment selection and motion planning. For low-level motion planning, prior world model-based reinforcement learning (RL) approaches have employed the similarity of latent states as a dense reward proxy. However, using latent similarity as an implicit distance metric yields inaccurate representations and depends heavily on predefined goal states. In this work, we address the limitations of similarity-based rewards by explicitly integrating 3D spatial information into the world model and deriving more accurate rewards from that distance-aware model for policy training. For high-level embodiment selection, prior model-based methods have either utilized reachability rewards inferred from predicted task rewards or relied on heuristic composite rewards. In this study, we retain the reachability reward and demonstrate that it yields superior performance in embodiment selection compared with composite reward formulations. By combining reachability-based selection and distance-guided planning, the model improves the performance of end-to-end learning for separate mobile manipulation behaviors. Extensive evaluations were conducted across two simulation environments with both fixed and randomized target settings. For motion planning, the distance reward achieves near 100% success rates across all configurations, significantly outperforming the baseline similarity reward, which achieves only 60% success in fixed-target scenarios and drops below 20% when targets are randomized. Regarding embodiment selection, the reachability-based separate control rapidly achieves a stable result with over 85% of arm activations occurring rationally within the near-goal region. Conversely, the heuristic composite reward slowly increases to a maximum of 60% for near-goal arm selections. These findings validate that the reachability reward provides a more effective and efficient mechanism for embodiment selection.