Differentiable Physics Training Method for Robot Motion Planning

Robotic manipulation in dynamic and unstructured environments remains a challenging problem, particularly for tasks that require precise interaction with moving targets. This work presents a robotic motion planning algorithm using differentiable simulation, where a robotic arm is trained to perform peg insertion tasks in an industrial setting where the insertion target is moving. Successfully performing such tasks requires real-time tracking of the target and continuous adaptation of the robot's trajectory. To achieve this, we train a neural network controller to generate actions along the trajectory, using differentiable physics to enable end-to-end optimization. The optimization objectives and non-penetration constraints are expressed in a differentiable way to perform gradient backpropagation through the entire simulation. Initial experimental results demonstrate that the proposed training approach is both fast and stable, yielding precise robot positioning and smooth trajectories, exceeding the results obtained with a more traditional Reinforcement Learning baseline.