Reproduce “Benchmarking Potential Based Rewards for Learning Humanoid Locomotion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Se Hwan Jeon, Steve Heim, Charles Khazoom, Sangbae Kim
This research empirically evaluates Potential-Based Reward Shaping (PBRS) against Direct Reward Shaping (DRS) in learning dynamic locomotion for an MIT Humanoid robot. The study demonstrates that PBRS offers significantly greater robustness to reward weight tuning and enhances training stability, achieving comparable or slightly improved performance over DRS.
Key insights
- PBRS provides substantially greater robustness to reward weight scaling compared to DRS, simplifying the reward design and tuning process in complex robotic applications.
- PBRS leads to significantly lower variance across multiple training runs, indicating enhanced learning stability and reliability in high-dimensional continuous control.
- Setting the discount factor to 1 within the PBRS potential function calculation was crucial for stable and faster convergence, pragmatically improving performance despite a theoretical deviation from strict policy invariance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.