Loco-manipulation using RL and Pretrained AMO Controller in MuJoCo

Combining reinforcement learning with a pretrained AMO (Any Motion) controller for loco-manipulation tasks in MuJoCo physics simulator.

This project explores long-horizon loco-manipulation for humanoids by building a hybrid control framework that integrates locomotion controllers with RL-based manipulation policies.

Overview

I am building a hybrid loco-manipulation control framework by integrating a locomotion controller with RL-based PPO-trained manipulation policies, enabling long-horizon factory tasks in MuJoCo with reward design. Additionally, I am training lightweight LLMs for SOP-driven high-level retrieval and planning to generate multi-step task plans and sequence MuJoCo skills with pre/post-condition checks.

Technical Approach

  • Simulation Environment: MuJoCo
  • Base Controller: Locomotion controller integrated with RL policies
  • Learning Method: Proximal Policy Optimization (PPO)
  • Task: Long-horizon loco-manipulation for humanoids
  • Additional Components: Lightweight LLMs for high-level planning and task sequencing
Robot performing loco-manipulation task in MuJoCo.

Results