Reproduce “LORD: Large Models based Opposite Reward Design for Autonomous Driving” 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
Xin Ye, Feng Tao, Abhirup Mallik, Burhaneddin Yaman, Liu Ren
Bosch Research North America and the Bosch Center for Artificial Intelligence developed LORD (Large Models based Opposite Reward Design), a framework for autonomous driving reinforcement learning that uses large pretrained models to define rewards based on avoiding undesired outcomes, leading to improved success rates and generalization in highway driving simulations.
Key insights
- Large pretrained models can more effectively understand and evaluate undesired outcomes (e.g., "collision") compared to abstract desired outcomes (e.g., "safe driving").
- Focusing rewards on avoiding specific negative outcomes provides clearer and more informative gradients for reinforcement learning, reducing ambiguity and reward hacking.
- The choice of observation representation (image, video, or text) impacts performance and generalization, with text-based observations showing particularly robust capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.