Authors: Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, Shuran Song (Columbia University, MIT, Toyota Research Institute) · arXiv: 2303.04137
Formulates robot visuomotor control as conditional denoising score matching over receding-horizon action sequences, naturally capturing multimodal human demonstration distributions without mode-averaging collapse.
Key Quantitative & Theoretical Takeaway: Predicting closed-loop action sequences of horizon T_p while executing T_a steps via receding-horizon control combines temporal action smoothness with real-time visual reactivity.
This paper introduces Diffusion Policy, a new way of generating robot behavior by representing a robots visuomotor policy as a conditional denoising diffusion process. We benchmark Diffusion Policy across 12 different tasks from 4 different robot manipulation benchmarks.
Standard MSE regression averages distinct valid trajectories (such as passing left vs right around an obstacle) into an invalid midpoint collision, whereas iterative Langevin denoising follows the score gradient into a single coherent trajectory mode.
Replace arxiv.org with nujan.app on any arXiv URL (for example, https://nujan.app/abs/2006.11239) to open the PDF and AI research partner side-by-side.