Robotic Manipulator Model Learning Without Velocity Sensors

The patent applies preliminary action by pre-training a simulation model using reinforcement learning before deploying it to the real physical system. The simulation environment allows the agent to learn optimal control policies through random exploration without risking damage to the actual robotic system. Once trained in simulation, the learned policy is transferred to the real system, eliminating the need for dangerous random exploration in the physical environment.

The patent applies the taking out principle by extracting and eliminating the requirement for velocity and acceleration sensors from the system. The method reformulates the dynamic model to depend only on position measurements, which are already available from standard encoder sensors. This extraction of the problematic measurement requirements simplifies the device while maintaining prediction accuracy through a restructured modeling approach that uses position-only data with augmented state variables.

The patent applies segmentation by dividing the computational task into distinct phases: offline training phase where the Gaussian Process model is trained using simulation data, and online execution phase where the trained model is used for prediction. The complex GP training computations are performed offline when computational resources are abundant, while the online phase requires minimal computation for real-time control decisions. This temporal segmentation reduces the immediate computational load during robot operation while maintaining high generalization capability.