Encoder-Adapted Sim-to-Real Transfer of Simulation-Trained Diffusion Policies for Robot Manipulation

In this paper, we propose simDP, a sim-to-real transfer framework that enables diffusion policies trained in simulations to be deployed on real-world robots. The key idea is to reduce the sim-to-real gap by aligning the action and observation spaces between simulation and reality. This design allows the action decoder trained in simulation to be reused in the real-world with minimal modification. We evaluated simDP on object manipulation tasks derived from the MimicGen benchmark and show that a simulation-trained diffusion decoder, when combined with a real-world adapted observation encoder, achieves task-completion performance similar to and in some cases better than diffusion policies trained only on limited real-world data. These results suggest that reusing a simulation-trained action decoder with lightweight real-world encoder adaptation provides a practical and scalable solution for sim-to-real transfer in embodied robotic learning.