Adapting to Variable Kinematic Configurations: A Causal-Kinematic Attention Approach for Robotic Control

Deep reinforcement learning (RL) policies for robotic control typically overfit to a single hardware configuration. Any change to the kinematic chain breaks the learned mapping and requires complete retraining. This paper investigates whether combining self-attention mechanisms with explicit spatial observations can improve a policy’s adaptability to varying kinematic topologies. We train a single RL agent to control a robotic manipulator across different degrees of freedom (DOF), ranging from a restricted 4-DOF mode to a fully redundant 7-DOF configuration. Instead of fixed-length state vectors, the method processes the active joints as a variable-length sequence in an attention buffer, enriching each joint’s representation with its relative spatial routing and geometric Jacobian influence. This structure allows the agent to dynamically evaluate the physical utility of its available actuators. The proposed Architecture reaches a success rate of 82.1% across trained topologies and 66.6% on unseen configurations, outperforming MLP and generic attention baselines while remaining nearly collisionfree. Finally, we demonstrate successful sim-to-real transfer by deploying the simulation-trained agent on a physical Franka Emika Panda manipulator.