Actuator-level DRL policy masters cable robots across configurations

Cable-driven parallel robots (CDPRs) are flexible but notoriously hard to control, and deep reinforcement learning (DRL) often struggles to generalize across different configurations or actuator counts. A team of researchers from CRAN, UIR, LARSEN, LORIA, and ISM has proposed a novel solution: instead of learning a policy that controls the whole robot to reach an end-effector position, they train an actuator-level policy (ALP) that independently commands each motor to achieve its target cable length. This shift means the policy is agnostic to robot geometry or how many motors are present, since it only learns the local mapping from cable length errors to motor actions.

Trained entirely in simulation on a simple 4-motor planar 2D CDPR, the ALP was successfully transferred to a real 8-motor CDPR performing full 3D motion—a leap that would typically require retraining from scratch. The authors report that this actuator-level approach outperforms traditional DRL methods on both robustness and precision. It also sidesteps the computationally challenging forward kinematics problem by relying on inverse kinematics instead. This is the first DRL work to control CDPRs at the actuator level, and it opens the door to 'train once, deploy anywhere' robotics control.