Reproduce “InDRiVE: Reward-Free World-Model Pretraining for Autonomous Driving via Latent Disagreement” 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
Feeza Khan Khanzada, Jaerock Kwon
The InDRiVE framework employs reward-free world-model pretraining for autonomous driving via latent ensemble disagreement, yielding robust zero-shot transfer and efficient few-shot adaptation for downstream tasks in a high-fidelity simulator. This approach achieved up to 94.8% success rate in unseen towns for collision avoidance after few-shot tuning, showing enhanced generalization compared to intrinsic and task-specific baselines.
Key insights
- Reward-free pretraining, driven by latent ensemble disagreement, can successfully produce a reusable driving world model and an effective driving behavior that transfers well to new tasks and unseen environments.
- Disagreement-based intrinsic motivation (InDRiVE) leads to more generalizable representations and policies compared to other intrinsic methods (ICM, RND) and task-specific training baselines in autonomous driving.
- The intrinsically pretrained world model substantially accelerates few-shot adaptation for safety-critical tasks, particularly collision avoidance, under significant distribution shifts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.