[论文解读] Towards Adaptive Environment Generation for Training Embodied Agents

Teresa Yeo, Dulaj Sanjaya Weerakoon+2 more|arXiv (Cornell University)|Feb 6, 2026

Multimodal Machine Learning Applications被引用 0

一句话总结

论文提出一个闭环框架，利用智能体轨迹分析自适应地修改训练环境，使 embodied navigation 任务能够逐渐变得更困难、以反馈驱动的课程设计。

ABSTRACT

Embodied agents struggle to generalize to new environments, even when those environments share similar underlying structures to their training settings. Most current approaches to generating these training environments follow an open-loop paradigm, without considering the agent's current performance. While procedural generation methods can produce diverse scenes, diversity without feedback from the agent is inefficient. The generated environments may be trivially easy, providing limited learning signal. To address this, we present a proof-of-concept for closed-loop environment generation that adapts difficulty to the agent's current capabilities. Our system employs a controllable environment representation, extracts fine-grained performance feedback beyond binary success or failure, and implements a closed-loop adaptation mechanism that translates this feedback into environment modifications. This feedback-driven approach generates training environments that more challenging in the ways the agent needs to improve, enabling more efficient learning and better generalization to novel settings.

研究动机与目标

• 为具身智能体在未见环境中的泛化能力提供更强的需求动机。
• 提出一个闭环系统，根据智能体表现自适应调整环境难度。
• 利用结构化的环境表示和细粒度轨迹反馈实现目标导向的课程设计。
• 通过使用LLMs进行分析和修改来证明可行性（概念验证）。
• 强调局限性并概述未来在评估与扩展方面的方向。

提出的方法

• 将环境表示为结构化场景图（O、A、R），以实现可控修改。
• 使用分析模型 F（如 GPT-4.1-mini）从智能体轨迹中提取成功、中间关切和高层次修改建议。
• 使用生成器 G（如 GPT-4.1-mini）将 F 的分析转化为具体的环境编辑，同时确保有效性和可解性。
• 实现考虑碰撞的放置，以在不产生对象相交的情况下实现修改。
• 渲染更新后的环境并迭代该循环，生成渐进式课程。
• 可选地讨论用于生成配置增量的梯度法与基于模型（LLM）的方法之间的权衡。

Figure 1: Embodied navigation performance is sensitive to object perturbations. Top-down view of agent trajectories (yellow to orange path) for an object navigation task with the fridge as the target object. In the training environment (left), the agent successfully navigates to the target, while in

实验结果

研究问题

• RQ1 闭环、反馈驱动的环境生成循环是否能提高具身智能体的学习效率？
• RQ2 如何通过细粒度轨迹分析引导有意义且现实的环境修改？
• RQ3 在自适应生成中，基于LLM的编辑与梯度法环境增量预测之间的权衡是什么？
• RQ4 在连续编辑后，生成的环境仍然可解且物理上合理吗？

主要发现

• 一个概念验证的闭环管线是可行的，利用轨迹分析来引导环境扰动。
• 环境修改聚焦于创造更具挑战性且更真实的场景（如更窄的通道），而非随机扰动。
• 可以将碰撞感知和可行性约束整合进来，以维持可行的环境。
• 该方法展示了自适应环境设计提升效率和泛化能力的潜力， though 全面的评估仍为未来工作。

Figure 2: Overview of the proposed adaptive environment generation framework. Starting from an original environment $e_{t}$ , an agent with policy $\pi_{t}$ is deployed to perform embodied tasks (e.g., object navigation), producing a top-down trajectory visualization $\tau^{e_{t}}$ . An analysis mod