Robotics & Embodied AI
Diffusion Policy outperformed comparison methods by 46.9 percent across 15 robot manipulation tasks
The authors report that Diffusion Policy, a way of choosing robot actions through a diffusion process, outperformed the methods they compared it with by an average of 46.9 percent across 15 tasks from four robot manipulation benchmarks. A visuomotor policy is the part of a robot system that turns camera observations into motor actions. Diffusion Policy treats that policy as a conditional denoising diffusion process, starting from noisy candidate actions and refining them step by step.
What did the study measure?
Researchers from Columbia University, Toyota Research Institute, the Moscow Institute of Thermal Technology, and Stanford University benchmarked Diffusion Policy on 15 tasks from four robot manipulation benchmarks. The abstract reports a consistent improvement over the comparison methods, with an average improvement of 46.9 percent. The abstract does not give individual task scores.
How does the method work?
Diffusion Policy learns the gradient of the action-distribution score function. During action selection, it repeatedly updates the candidate action through stochastic Langevin dynamics steps. The paper also adds receding horizon control, which plans a short sequence of actions ahead, visual conditioning, and a time-series diffusion transformer.
What did the authors say makes the approach useful?
The authors report that the diffusion form handles multimodal action distributions. They also report that it suits high-dimensional action spaces and shows impressive training stability.
Frequently asked questions
Is the 46.9 percent gain a single benchmark score?
The abstract says the figure is an average improvement across the 15 tasks from the four benchmarks. It does not break out scores for individual tasks.
Did the authors test on physical robots?
The abstract does not state that the 46.9 percent result was measured on physical robots. It describes technical contributions for physical robots, but the benchmark wording leaves the setting of the measured result unstated.
Who funded the work?
The abstract does not list funding sources. One listed author affiliation is the Moscow Institute of Thermal Technology.
Limits and what's next
The abstract does not state limitations such as failure cases, safety risks, or how well the result transfers outside the four benchmarks. It is a peer-reviewed article, published in The International Journal of Robotics Research on October 11, 2024. The authors say they hope the work will motivate new policy learning techniques that use diffusion models.
Source: diffusion-policy.cs.columbia.edu