Rose-Hulman Institute of Technology Ongoing research

Humanoid Locomotion Research

Reinforcement Learning and Sim-to-Real Transfer

Learning locomotion that transfers

I am investigating how bipedal locomotion policies are trained and evaluated, and how they transfer between simulation and physical platforms. Berkeley Humanoid Lite is an open-source physical platform being established for locomotion research at Rose-Hulman.

The platform is supporting infrastructure for a broader research goal: reproducible experiments in learning-based control.

Research questions

  • How do velocity-command curricula influence command tracking and stability?
  • How robustly do locomotion policies transfer across simulators and physical systems?

Platform & methods

Training workflow

An Isaac Lab / Isaac Sim reinforcement learning workflow supports policy training investigations.

Cross-simulator evaluation

MuJoCo provides a setting for sim-to-sim policy evaluation.

Curriculum and tracking

Current interests include velocity-command curricula, command tracking, and robust policy evaluation.

Detailed observation, action, reward, and curriculum specifications will be added with verified experiment documentation.

Experiments & evaluation

The evaluation goals are to study tracking, stability, and robustness across training conditions and simulators. Measured comparisons and experiment media have not yet been supplied for this portfolio.

Evaluation figures & demonstrations coming soon

No measured improvement or successful physical deployment is claimed here.

Transfer challenges & next steps

Research interests include understanding simulation mismatch, evaluating policy robustness, and improving velocity-command curricula. A further goal is to document a reproducible workflow for future student researchers.

Hardware validation remains to be documented; simulation results and physical results will be identified separately.

Research documents & code

Report coming soon

The compiled research PDF will be available here, with direct open and download links.

Code and demonstration links will be added when public resources are available.