Learning and Planning in the Wild: Neuro-Symbolic Agents for Data-Efficient and Interpretable Robot Autonomy

Deploying robots in complex, dynamic environments requires systems that can learn efficiently from humans, adapt autonomously to unforeseen changes, and operate within real-world resource constraints. Current approaches face a fundamental tension: purely data-driven methods require prohibitive amounts of data and computation, while classical symbolic systems are brittle and cannot recover from novelty. This dissertation resolves that tension through a unified neuro-symbolic framework that bridges continuous neural learning with discrete symbolic reasoning, enabling robots to acquire complex skills from minimal input and generalize robustly to novel situations. The framework makes six core contributions toward establishing neuro-symbolic integration as a scalable and interpretable paradigm for open-world robotics. First, it extends the open-world adaptation challenge from discrete toy environments to complex continuous robotics domains, grounding the problem in realistic manipulation settings. Second, a bi-level planning and learning architecture forges connections between symbolic reasoning and neural training, enabling faster adaptation to unexpected events by conditioning and anchoring existing policies with known symbols. Third, a curiosity-driven adaptation architecture deepens this coupling further by integrating intrinsic curiosity at the action level and symbolic world models at the planning level, making them co-evolve: action-level exploration discovers new behavioral traces that are captured symbolically, while symbolic knowledge guides exploitation toward goal completion via reward machines, yielding substantially faster detection and recovery from runtime novelty than state-of-the-art hybrid baselines. Fourth, a hierarchical and nested task and motion planning framework paired with modelbased reinforcement learning exploits the composability inherent in neurosymbolic architectures to tackle long-horizon structured tasks, enabling zero-shot adaptation to unseen environmental conditions without retraining. Fifth, a few-shot imitation learning system extracts both low-level control policies and symbolic domain knowledge from as few as five demonstrations, enabling strong generalization across six robotic domains including a real automated forklift for industrial manipulation. Sixth, empirical evaluation shows that this neuro-symbolic approach achieves high task success, robust generalization to unseen configurations, and greatly lower energy and time costs compared to fine-tuned Vision-Language-Action models. Taken together, these contributions demonstrate that tight, bidirectional integration between symbolic and neural representations, offers a principled path toward autonomous systems capable of reliably achieving complex long-horizon tasks across diverse and unpredictable real-world environments.