Adaptive approximation of dynamics gradients via interpolation to speed up trajectory optimisation

Authors: David Russell, Rafael Papallas and Mehmet Dogar
Conference: IEEE International Conference on Robotics and Automation (ICRA) 2023

Our work proposes a method of speeding up gradient-based trajectory optimisation, specifically in tasks where contact is involved, and analytical dynamics gradients can not be constructed. In contact-based tasks, numerically approximating gradients via methods like finite-differencing is a widely used method. This method is very time consuming when querying expensive physics simulators and grows more computationally expensive as the complexity of the model increases. We propose computing the gradients using finite-differencing at “key-points” over the trajectory and then interpolating cheap approximations to the gradients in between. We show that this methodology can speed up trajectory optimisation significantly whilst retaining an almost identical quality of final solution. Decreasing the time taken for trajectory optimisation is important in industry applications as currently trajectory optimisation for manipulation is slow when the scenes become cluttered with obstacles.