Real-time Multi-Objective Trajectory Optimization

In this article, a method is presented for generating a trajectory through predetermined waypoints, with jerk and duration as two conflicting objectives. The method uses a Seq2Seq neural network model to approximate Pareto efficient solutions. It trains on a set of random trajectories optimized by Sequential Quadratic Programming (SQP) with a novel initialization strategy. We consider an example pick-and-place task for a robot manipulator. Based on several metrics, we show that our model generalizes over diverse paths, outperforms a genetic algorithm, SQP with naive initialization, and scaled time-optimal methods. At the same time, our model features a negligible GPU-accelerated inference time of 5ms that demonstrates applicability of the approach for real-time control.