This was a team research project at the UM-SJTU Joint Institute, from February to June 2024. The team explored using an LLM to propose robotic manipulation tasks, then generate the scenes and supervision needed for training.
Contributions
Set up simulation environments in Robosuite and expanded the training data.
Ran imitation-learning and Soft Actor-Critic experiments, including learning from low-quality demonstrations.
Studied DDPM and Diffusion Policy for motion prediction and planning.