📄 中文摘要
自主系统在开放世界环境中适应不可预见的新颖性是一个重大挑战。混合规划和强化学习（RL）方法虽然前景广阔，但通常存在样本效率低下、适应缓慢和灾难性遗忘的问题。为解决这些局限性，本框架提出一种神经符号学习方法，该方法将分层抽象、任务与运动规划（TAMP）以及强化学习相结合，以实现机器人的快速适应。该方法通过结合符号推理的结构化知识与深度学习的感知能力，使机器人能够在面对未知情况时，快速理解并执行复杂任务。具体而言，分层抽象将高层任务分解为可管理的子任务，并通过符号表示进行建模，从而提供任务执行的逻辑结构。TAMP模块利用这些符号表示来生成可行的动作序列和运动轨迹，确保机器人能够物理上实现规划的动作。强化学习组件则负责在实际执行过程中，通过与环境的交互来优化低层控制策略和适应未知动态。这种集成方式允许系统在遇到任务僵局（例如，规划器无法找到解决方案或执行器遇到意外障碍）时，能够快速识别并生成新的解决方案。通过神经符号融合，系统能够从少量经验中学习新的任务范式，并有效避免灾难性遗忘，因为符号知识提供了稳定的结构，而神经网络则负责适应和泛化。这种方法显著提升了机器人在开放世界场景中的鲁棒性、自主性和学习效率，使其能够更有效地处理复杂、动态且充满不确定性的现实世界任务。
📄 English Summary
Breaking Task Impasses Quickly: Adaptive Neuro-Symbolic Learning for Open-World Robotics
Adapting to unforeseen novelties in open-world environments remains a major challenge for autonomous systems. While hybrid planning and reinforcement learning (RL) approaches show promise, they often suffer from sample inefficiency, slow adaptation, and catastrophic forgetting. To address these limitations, a neuro-symbolic framework is proposed, integrating hierarchical abstractions, task and motion planning (TAMP), and reinforcement learning to enable rapid adaptation in robots. This approach combines the structured knowledge of symbolic reasoning with the perceptual capabilities of deep learning, allowing robots to quickly understand and execute complex tasks when facing unknown situations. Specifically, hierarchical abstractions decompose high-level tasks into manageable subtasks, modeled using symbolic representations, thereby providing a logical structure for task execution. The TAMP module leverages these symbolic representations to generate feasible action sequences and motion trajectories, ensuring the robot can physically realize the planned actions. The reinforcement learning component is responsible for optimizing low-level control policies and adapting to unknown dynamics through interaction with the environment during actual execution. This integrated approach allows the system to quickly identify and generate new solutions when encountering task impasses (e.g., a planner failing to find a solution or an executor encountering unexpected obstacles). Through neuro-symbolic fusion, the system can learn new task paradigms from limited experience and effectively avoid catastrophic forgetting, as symbolic knowledge provides a stable structure while neural networks handle adaptation and generalization. This method significantly enhances the robustness, autonomy, and learning efficiency of robots in open-world scenarios, enabling them to more effectively handle complex, dynamic, and uncertain real-world tasks.