Visuomotor Diffusion Policy & Dexterous Manipulation Lab
An interactive engineering laboratory for physical robot manipulation: simulate DDIM trajectory denoising over continuous multi-step action chunks, photometric stereo tactile contact arrays (GelSight/DIGIT), and multi-rate impedance control loops.
The Architecture of Visuomotor Diffusion Policies
Traditional reinforcement learning and behavioral cloning struggle with multimodal action distributions (e.g. going left versus right around an obstacle). Diffusion policies express complex multimodal distributions by formulating robot action generation as a conditional denoising diffusion process over continuous trajectories.
1. Action Chunking & Temporal Horizon
Instead of predicting single-step actions at each frame, the policy outputs a sequence chunk A = [a_t, a_{t+1}, ..., a_{t+k}]. Recency-weighted temporal ensembling smooths consecutive overlapping predictions, eliminating jerky joint oscillations.
2. Score-Based Trajectory Denoising
Conditioned on visual tokens from head and wrist cameras, a Denoising Diffusion Implicit Model (DDIM) iteratively denoises a random Gaussian trajectory vector into smooth 6-DoF end-effector waypoints and gripper aperture targets in 10 fast steps.
3. High-Frequency Tactile Adaptation
While vision tokens update at 10–20 Hz, elastomeric tactile sensors (GelSight/DIGIT) detect incipient micro-slip at 60–100 Hz, allowing the low-level impedance controller to dynamically adjust normal squeeze forces before objects drop.