Dual-Arm Neural Motion Planning Using Action Chunking Transformers

Deploying humanoid robots in everyday environments requires rapidly generating coordinated and collisionfree motions for two arms. Despite recent advances in global motion planning and trajectory optimization, high-dimensional configuration spaces still pose a significant challenge and can cause long planning delays. Neural motion planning addresses this shortcoming by leveraging large planning problem datasets to train a motion planning policy. Such a model can then be used online or to seed global planners. In this work, we extend the neural motion planning approach to a dual-arm setting and thus generate motions in a $\mathbf{1 4}$-dimensional joint space. The proposed method is based on a transformer architecture and combines proprioceptive and geometric information in the form of point clouds to predict chunks of joint-space actions. In contrast to previous approaches, we use recent advances in imitation learning for fast point cloud tokenization. We show the performance of our model in simulation and provide insights for deploying it on a real robot. In our experiments, we use the model as motion planner and for seeding a state-of-the-art trajectory optimizer. Our results show the advantage of transformer-based neural motion planning and its positive impact on the success rates of traditional solvers.