Quadruped RL with 2-DoF Spine

Current month-long challenge: hardware modification plus reinforcement learning control.

$cat case-study.md

Summary

Feb 2026 - Present | Personal challenge | Tempe, AZ

Modified a Unitree Go2 by introducing a 2-joint spine and trained an RL policy for narrow-hallway, tight-corner navigation.

What I built

  • Mechanical change to add 2-DoF spine behavior.
  • Policy training loop in Isaac Sim.
  • Observation and action design using IMU + joint states + relative obstacle/goal pose.
  • Iterative environment and reward tuning for robust cornering.

Current status

  • Policy can complete hallway turns in nominal simulation conditions.
  • Current focus is robustness under dynamics variation.
  • Final challenge metrics to be added after completion.

github back-to-projects demo (todo)