Abstract
The present disclosure provides a method of learning physical interactions in robotic manipulation, comprising formulating a learning representation that includes a spatial force of a robotic manipulator, a spatial stiffness, and a spatial damping of the manipulator, wherein the learning representation is used by an artificial intelligence agent to interpret or learn the dynamic interaction between the robotic manipulator and an object or environment. The spatial stiffness and spatial damping are derived from joint-space stiffness and damping parameters using Jacobian-based transformations. The learning representation may be provided in forms comprising spatial force vectors, stiffness matrices, damping matrices, end effector pose data, gripper force measurements, and vision sensor data. The representation serves as input to artificial intelligence learning algorithms including reinforcement learning, supervised learning, or imitation learning, enabling the agent to learn control policies for manipulating objects or regulating contact forces.