Robot Control With Haptic Reinforcement Learning for Contact-Rich Assembly

The patent implements reinforcement learning with haptic feedback from force/torque sensors, where the robot continuously receives feedback about contact forces during assembly tasks and adjusts its control policy accordingly. This enables the system to learn from interaction rather than relying on pre-programmed trial-and-error processes, significantly reducing training time while maintaining control accuracy.

The robot autonomously learns assembly skills through self-directed interaction with the environment using reinforcement learning. The system independently optimizes its control policy by processing haptic feedback and updating its neural network parameters, eliminating the need for extensive manual programming or supervised trial-and-error procedures.

The patent introduces haptic feedback from force/torque sensors as an intermediary that bridges the robot's actions and environmental responses. This tactile information serves as a rich feedback signal that enables the reinforcement learning system to learn complex assembly skills without requiring overly complicated system architectures, as the haptic feedback encapsulates essential interaction dynamics.

The patent replaces traditional mechanical control systems with reinforcement learning-based control. Instead of using complex mechanical mechanisms and explicit material models, the system uses deep neural networks that learn optimal control policies directly from haptic feedback, simplifying the overall system architecture while enhancing adaptability.