Simulated Robotic Soft Body Manipulation

The performance of intelligent agents manipulating a soft body object depends on the agent's understanding of the execution environment. Hence, by keeping the agent fixed and changing the environment, the difference between environments can be measured. However, this becomes complicated when dealing with agents that learn in each environment. We propose a framework for evaluating the influence of the simulated soft bodies (and related object models) on the reinforcement learning algorithms' performance. The change in algorithm behavior is quantified between different environments, and the correlation of behavioral difference is measured via statistical analysis. An evaluation case is presented on PyBullet and MuJoCo physics simulation environments with DDPG, PPO, TD3 and SAC algorithms.