Neuro-Symbolic Learning for Long-Horizon Task Planning Under Complex Logical Constraints

Researchers present a neuro-symbolic learning framework that addresses a critical inefficiency in robotic task planning by combining neural networks with symbolic planning under complex logical constraints. The method uses bilevel optimization to learn object-importance scores while solving planning problems in pruned search spaces, reducing planning failures by 80% and planning time by 57% across multiple benchmarks and real-world robotic applications.

This research tackles a fundamental problem in autonomous robotics: efficiently planning long-horizon tasks when robots must reason about object affordances, spatial relationships, and action dependencies. Traditional neuro-symbolic approaches train on complete search spaces but deploy in pruned spaces defined by their own predictions, creating a distribution mismatch that degrades real-world performance. The proposed bilevel optimization framework elegantly addresses this exposure bias by making the learning process mirror deployment conditions, where the neural scorer's predictions directly constrain the symbolic planner's search space.

The 3R recovery strategy—parallel Repair, Restart, and Rollback mechanisms—provides the critical stabilization needed for this bidirectional learning process, enabling the upper-level neural network to learn from more realistic feedback rather than oracle labels. This represents a meaningful methodological advance in bridging the neural-symbolic gap. The results demonstrate substantial practical improvements: an 80% reduction in planning failures and 57% reduction in planning time position this work as a significant step toward deployable autonomous systems.

For the robotics and AI industries, this framework has immediate implications for mobile manipulators and other complex robotic systems that struggle with combinatorial planning. The successful validation on both simulated and real quadruped-based systems suggests the approach generalizes beyond toy problems. As robots increasingly operate in unstructured environments with complex constraints, more efficient planning methods directly translate to increased deployment feasibility and reduced computational overhead, making this work valuable for companies developing autonomous manipulation and navigation systems.

• →Bilevel optimization framework addresses the train-test mismatch in neuro-symbolic planning by learning within pruned search spaces that mirror deployment conditions.
• →The 3R recovery strategy (Repair, Restart, Rollback) stabilizes learning by providing reliable feedback from realistic planning scenarios.
• →Framework achieves 80% reduction in planning failures and 57% reduction in planning time across three benchmarks.
• →Real-world validation on mobile manipulators demonstrates practical viability beyond academic benchmarks.
• →Approach scales to long-horizon tasks with complex logical constraints, addressing a major bottleneck in autonomous robotics.