Functional Locomotion Policy for rough terrain traversal for Wheeled Biped
Wheeled Biped
Reinforcement Learning
Isaac Sim
Locomotion
Research Objective
Develop a functional locomotion policy for a wheeled biped that can traverse rough warehouse terrain while transporting liquids with minimal spillage. The objective is to achieve stable, safe, and efficient motion under disturbances and uneven surfaces, targeting practical workflows in industrial settings.
Key Features Achieved
- Established a stable standing policy that reliably holds an equilibrium point under small perturbations.
- Progressed velocity-control behaviors on flat terrain with prioritized tracking and safe action bounds.
- Defined training scaffolds and diagnostics to iteratively expand from standing to commanded motion.
- Work in progress: extending to rough-terrain traversal and slosh-aware motion objectives.
Challenges & Learnings
- Rapidly evolving simulation stack (Isaac Lab vs. Isaac Sim) required frequent re-instrumentation and environment parity checks.
- Bridging conventional control intuition with RL required careful reward shaping, observation design, and safety envelopes for a generalized policy.
- Iterative domain setups and curriculum design improved stability and sample efficiency across tasks.