外观
论文arXiv 2610.0072928 天前
奖励作为观测：学习基于奖励的策略以实现快速适应
原文链接：Reward as Observation: Learning Reward-Based Policies for Rapid Adaptation（arXiv 2610.00729 · 发布日期：2026-09-30）
Reward as Observation: Learning Reward-Based Policies for Rapid Adaptation
针对观测空间完全不同环境间迁移，提出仅以奖励和动作为条件的策略，并在三类环境及Habitat-Sim导航中验证零样本适应。
核心要点
- 深度神经网络策略在新环境适应中常需大量样本，难以像人类一样快速迁移。
- 方法将策略仅条件化于奖励和动作，讨论其挑战与可行性，并提出实用训练算法。
- 在Pointmass、Cartpole和2D Car Racing中训练后，可零样本迁移到不同观测形式。
- 奖励策略还可在目标环境中进一步指导基于观测的策略训练。
验证范围与局限：验证范围为Pointmass、Cartpole、2D Car Racing及Habitat-Sim中的Stretch机器人导航。
英文摘要 (Abstract)
This paper explores a reward-based policy to achieve zero-shot transfer between source and target environments with completely different observation spaces. While humans can demonstrate impressive adaptation capabilities, deep neural network policies often struggle to adapt to a new environment and require a considerable amount of samples for successful transfer. Instead, we propose a novel reward-based policy only conditioned on rewards and actions, enabling zero-shot adaptation to new environments with completely different observations. We discuss the challenges and feasibility of a reward-based policy and then propose a practical algorithm for training. We demonstrate that a reward policy can be trained within three different environments, Pointmass, Cartpole, and 2D Car Racing, and transferred to completely different observations, such as different color palettes or 3D rendering, or Stretch robot navigation in Habitat-Sim, in a zero-shot manner. We also demonstrate that a reward-based policy can further guide the training of an observation-based policy in the target environment.
作者：Morgan Byrd, Jacob Blevins, Maks Sorokin, Robert Wright, Sehoon Ha
来源
本页事实对应的公开页面，以原文为准。