Reinforcement Learning for Whole Body Manipulation

Building on top of the MuJoCo simulation I’ve built, I’m working on training Baloo to manipulate large objects:

Each object has a random size and mass and the objective is to lift the object off the floor successfully. I’m using the implementation of PPO from the Stable Baselines3 library.

For this project, I’m focused on a few different aspects:

• Reward shaping: The reward function is crucial for training the robot to perform the task. I’m experimenting with different reward functions to see which one works best.
• Imitation learning: I’m exploring how an expert trajectory can help accelerate learning.
• Sim2real: I’m working on transferring the learned policy to the real robot. This is a challenging problem, but I’ve put in lots of effort to maximize the realism of the MuJoCo simulation.
• Tactile Sensing: I’m exploring the effects of whole-body tactile sensing on the learning process.