The study focuses on dynamic motion generation tasks, such as contact and collisions, where small changes in policy parameters can have a significant impact on the outcomes. For instance, in soccer, even a slight variation in the hitting position, applied force, or ball friction can result in the ball flying in completely different directions. However, it is unlikely that completely different skills are required for heading the ball in different directions.
To address this issue, the researchers proposed a multitask reinforcement learning algorithm. This algorithm aims to adapt the policy to implicit changes in goals or environments within a single motion category. These changes can include different reward functions or physical parameters of the environment.
To evaluate their proposed method, the researchers used a monopod robot model in a ball heading task. The results demonstrated that the proposed method successfully adapted to implicit changes in goal positions or ball coefficients of restitution. On the other hand, the standard domain randomization approach was unable to handle the different task settings.