Keywords: Robot Learning, Imitation Learning, 3D Scene Representation
Abstract: Imitation learning has shown remarkable capability in executing complex robotic manipulation tasks. However, existing frameworks often fall short in structured modeling of the environment, lacking explicit characterization of geometry and semantics, which limits their ability to generalize to unseen objects and layouts. To enhance the generalization capabilities of imitation learning agents, we introduce a novel framework in this work, incorporating explicit spatial and semantic information via 3D semantic fields. We begin by generating 3D descriptor fields from multi-view RGBD observations with the help of large foundational vision models. These high-dimensional descriptor fields are then converted into low-dimensional semantic fields, which aids in the efficient training of a diffusion-based imitation learning policy. The proposed method offers explicit consideration of geometry and semantics, enabling strong generalization capabilities in tasks that require category-level generalization, resolving geometric ambiguities, and attention to subtle geometric details. We evaluate our method across eight tasks involving articulated objects and instances with varying shapes and textures from multiple object categories. Our method proves its effectiveness by outperforming state-of-the-art imitation learning baselines on unseen testing instances by 57%. Additionally, we provide a detailed analysis and visualization to interpret the sources of performance gain and explain how our method can generalize to novel instances.
Submission Number: 5
Loading