xTED: Cross-Domain Adaptation via Diffusion-Based Trajectory Editing

Reusing cross-domain data is vital for decision-making tasks with limited target data, yet existing methods often rely on complex, inflexible model-level designs. To address this, we propose the Cross-Domain Trajectory EDiting (xTED) framework, which bridges domain gaps directly at the data level. Our proposed model architecture effectively captures the intricate dependencies among states, actions, and rewards, as well as the dynamics patterns within target data. Edited by adding noises and denoising with the pre-trained target diffusion model, source domain trajectories can be transformed to align with target domain properties while preserving original task semantic information. This process effectively corrects underlying domain gaps, enhancing state realism and dynamics reliability in source data, and allowing flexible integration with various single-domain and cross-domain downstream policy learning methods. Despite its simplicity, xTED demonstrates superior performance in extensive simulation and https://xted24.github.io/xTED/ real-robot experiments. The full version can be found at https://arxiv.org/abs/2409.08687.