simDP: Sim-to-Real Transfer with Shared Action Spaces.

In this paper, we propose simDP, a sim-to-real transfer framework that enables diffusion policies trained in simulations for the efficient deployment 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. Specifically, we reformulate the action space using end-effector pose and binary gripper state, which can be shared between simulated and physical robots. In addition, we use camera-based visual observations as the primary sensing modality in both domains and train a real-world observation encoder to align with the latent representation learned in simulation. 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, obtained across four manipulation tasks in a calibrated real-world transfer setting, suggest that reusing a simulation-trained action decoder with lightweight real-world encoder adaptation provides an effective strategy for controlled sim-to-real transfer, while broader evaluation across diverse tasks, environments, and robot embodiments remains an important direction for future work.