Overview

Contributed to Lwlab, a large-scale robotics simulation framework for daily-life task training, focusing on reinforcement learning pipeline migration and humanoid robot loco-manipulation capabilities.

• Pipeline Migration: Migrated reinforcement learning pipelines from Isaac Lab to Lwlab’s flexible configuration system
• Multi-robot Support: Enabled support for diverse robots including Untree G1/PandaOmron
• Teleoperation: Integrated VR/keyboard teleoperation capabilities
• Scene Customization: Implemented custom scene layout support

Loco-manipulation Improvements

• Velocity Tracking: Developed velocity-tracking policies for bipedal locomotion
• Control Decoupling: Implemented upper/lower-body control decoupling
• Gait Optimization: Optimized gait stability through reward-shaping and symmetric/smoothness losses