Robot learning preprint reports gains across multiple manipulation tasks

Researchers report that SR-DP, a method for training one robot control model across multiple tasks, achieved a 68.4% average success rate on six RoboTwin tasks. In an arXiv preprint, they report gains of 21.8 points over multi-task VO-DP and 3.8 points over the strongest per-task baseline. The method combines a frozen visual encoder with a training approach that routes updates through separate mathematical components of the control model. It targets interference between incompatible actions when tasks share a model. On four real-world tasks requiring two robot arms, the second training stage achieved a 43.0% average success rate, improving on the first stage by 19.0 points.