paper_rob__3d_diffusion_policy

What problem does it solve?

DP3 enables learning and inference of visuomotor actions from sparse 3D point clouds for robotics, reducing reliance on dense sensor data or extensive demonstrations.

Core Features & Use Cases

• DP3 policy overview: A diffusion-based policy that conditions action generation on compact 3D features extracted from sparse point clouds and proprioceptive state.
• Encoder & diffusion architecture: PointNet-based encoder (DP3Encoder) combined with a FiLM-conditioned 1D UNet to produce denoised action sequences.
• Use Case: Learn visuomotor policies for tasks like object manipulation with few demonstrations in both simulation and real robots.

Quick Start

Provide a point_cloud and agent_pos observation to the DP3 policy to generate a sequence of actions.