Multi-objective optimization-learned vs. hand-tuned task controllers on Talos robot

Task priority-based control weights and gains are often time-consuming to hand-tune, and because of this it is typical to only produce a single hand-tuned set of control parameters. This single controller often compromises on accuracy since it must work well for many different types of robot trajectories. We propose to solve this issue by learning a Pareto front of possible controllers using a multi-objective optimization algorithm. This video shows the performance of solutions learned with this approach compared to that of hand-tuned solutions, on the Talos humanoid robot.