Adapting to Variable Active-Joint Configurations via Kinematics-Informed Attention for Robot Arms

Deep reinforcement learning (RL) policies for robotic control are typically trained for a single hardware configuration. When a joint locks, the learned mapping no longer matches the available actuation. This paper investigates whether combining self-attention with explicit kinematic observations can improve a policy’s adaptability to varying active-joint configurations. We train a single RL agent to control a Franka Emika Panda manipulator across different sets of active joints, from a restricted 4-degree-of-freedom (DOF) mode to the fully redundant 7-DOF configuration. Instead of fixed-length state vectors, the method processes the active joints as a variable-length sequence, enriching each joint’s representation with its relative spatial offset and geometric Jacobian influence. The proposed kinematics-informed attention architecture reaches a success rate of 82.1% across trained configurations and 66.6% on unseen configurations, outperforming multilayer perceptron and generic attention baselines while remaining nearly collision-free. Deployed on a physical Panda, the simulation-trained agent reaches 79.0% success on the unseen configurations, closely matching the simulation on the same target schedule (77.0%).