3D Diffuser Actor: Multi-task 3D Robot Manipulation with Iterative Error Feedback
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Diffusion, Robot Learning, 3D Representations
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: We present 3D Diffuser Actor, a framework that marries diffusion policies and 3D scene representations for robot manipulation. Diffusion policies capture the action distribution conditioned on the robot and environment state using conditional diffusion models. They have recently shown to outperform both deterministic and alternative generative policy formulations in learning from demonstrations. 3D robot policies use 3D scene feature representations aggregated from single or multiple 2D image views using sensed depth. They typically generalize better than their 2D counterparts in novel viewpoints. We unify these two lines of work and present a neural policy architecture that uses 3D scene representations to iteratively denoise robot 3D rotations and translations for a given language task description. At each denoising iteration, our model “grounds" the current end-effector estimate in the 3D scene workspace, featurizes it using 3D relative position attentions and predicts its 3D translation and rotation error. We test 3D Diffuser Actor on learning from demonstrations in simulation and in the real world. We show our model outperforms both 3D policies and 2D diffusion policies and sets a new state of the art on RLBench, an established learning from demonstrations benchmark, where it outperforms the previous SOTA with a 12% absolute gain. We ablate our architectural design choices, such as translation invariance through 3D grounding and relative 3D transformers, and show they help model generalization. Our results suggest that 3D scene representations and powerful generative modeling are key to efficient learning of multi-task robot policies.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3846
Loading