Researchers led by Lixin Xu have trained DexFormer 2.0, a single reinforcement-learning policy for dexterous grasping that works across many combinations of robot arms and hands, according to a paper on arXiv. Grasping policies are usually trained for one arm with one hand, so the number of policies grows with every new pairing, the authors write.
The policy maps every robot to a common joint interface, encodes each joint's layout with attention shaped by the robot's kinematic tree and uses a memory to infer each robot's dynamics from recent interaction. After training on 18 pairings of four arms and six hands, it achieved 89% stable success on those pairings and 51% on six unseen ones, beating GRU and DexFormer 1.0 baselines.
Without fine-tuning, it also grasped with four new types of hand and, through retargeting, with parallel grippers. On a real Flexiv arm, student policies distilled from it worked straight from simulation with two hands, including one pairing it had not seen.



