TrueRMA: Learning Fast and Smooth Robot Trajectories with Recursive Midpoint Adaptations in Cartesian Space

We present TrueRMA, a data-efficient, model-free method to learn\ncost-optimized robot trajectories over a wide range of starting points and\nendpoints. The key idea is to calculate trajectory waypoints in Cartesian space\nby recursively predicting orthogonal adaptations relative to the midpoints of\nstraight lines. We generate a differentiable path by adding circular blends\naround the waypoints, calculate the corresponding joint positions with an\ninverse kinematics solver and calculate a time-optimal parameterization\nconsidering velocity and acceleration limits. During training, the trajectory\nis executed in a physics simulator and costs are assigned according to a\nuser-specified cost function which is not required to be differentiable. Given\na starting point and an endpoint as input, a neural network is trained to\npredict midpoint adaptations that minimize the cost of the resulting trajectory\nvia reinforcement learning. We successfully train a KUKA iiwa robot to keep a\nball on a plate while moving between specified points and compare the\nperformance of TrueRMA against two baselines. The results show that our method\nrequires less training data to learn the task while generating shorter and\nfaster trajectories.\n