Robotic systems operating in dynamic environments often struggle to adapt their movements to external sensory signals without relying on explicit analytical models or conventional position-tracking control schemes. This limitation is especially problematic in systems where detailed dynamic models or positional feedback are limited or unreliable. Addressing this gap requires control architectures that, like human sensorimotor systems, can learn and generalize from experience without requiring an accurate model of the system's dynamics. In this paper, we present a human-inspired controller based on sensorimotor learning for articulated robotic arms. Departing from conventional control strategies relying on precise dynamic models, this approach emphasizes adaptability to dynamic tasks and external stimuli while accounting for gravity and system dynamics. The controller operates in two phases: an offline exploration phase that builds a discretized database of system responses for sampled initial states and actuator current commands, and an online exploitation phase that selects control actions by matching the current state to the closest explored state and applying a probabilistic command-selection rule. Rather than relying on an explicit analytical dynamics model, the approach uses empirical state–action–response relationships learned from the system's own behavior. During exploitation, the current joint positions and velocities are used to identify the closest explored state, while the command selection itself is based on velocity-response distributions learned during exploration. The controller's evaluation on various dynamic trajectories, including sinusoidal and trapezoidal profiles, demonstrates the feasibility of the framework on a 2-DOF RR robotic arm. Further analysis investigates the influence of control time parameters, the impact of actuator friction, and the controller's redeployability to modified mechanical configurations after re-exploration. These results underline the potential of biologically inspired learning mechanisms for data-driven robotic control while also highlighting limitations related to discretization, scalability, and future online adaptation.
Marchal et al. (Fri,) studied this question.