A preprint proposes Spectrally Routed Diffusion Policy (SR-DP) for robots that must learn several manipulation tasks in a single model. It addresses two problems: weak visual features and training signals from different tasks interfering with each other.
On six RoboTwin simulation tasks it averages 68.4% success, 21.8 points above multi-task VO-DP. On four real two-armed tasks, the second stage reaches 43.0% average success, 19.0 points above the first stage.
Real-robot success is below half, and the comparison is the authors' own.
Source: arxiv.org
Limits
Authors' own comparison; real-robot average 43.0% on four tasks.