Researchers have developed ELVIS, a novel approach to long-horizon visual planning in reinforcement learning that uses a Gaussian-mixture model predictive controller to maintain multiple hypotheses over extended rollouts. This method, detailed in a new paper, also incorporates an uncertainty-aware return mechanism to stabilize imagination and limit compounding errors. ELVIS demonstrates state-of-the-art performance on visual control tasks and shows promise for real-world applications with occlusions. Separately, another paper introduces TRAP, a backdoor attack targeting world models by manipulating the ranking of imagined trajectories, which has shown to degrade performance on agents like DreamerV3 and TD-MPC2. AI
影响 New methods for long-horizon planning and security evaluations for world models could advance agent capabilities and safety.
排序理由 Two new arXiv papers detail advancements in reinforcement learning planning and introduce a novel attack vector against world models.
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
