Neurosymbolic Embodied Agents

2026-08-17

AI Summary

This study addresses the inexecutability and environmental inconsistency of language model-based embodied planning by proposing a neuro-symbolic agent. The approach decomposes long-horizon tasks into visual exploration and constrained symbolic planning, leveraging PDDL-constrained decoding integrated with Monte Carlo Tree Search to ensure construct-level executability without additional training. Our analysis confirms that symbolic constraints and search are complementary rather than interchangeable. Experiments demonstrate that models ranging from 4B to 27B parameters achieve over 90% success rates in household tasks. Notably, smaller models outperform larger counterparts while significantly reducing token consumption, thereby validating the method’s efficiency and practicality for embodied AI applications.

Abstract

Language and vision-language models generate plausible embodied plans but do not guarantee executability, as their outputs can violate environment dynamics or act on incorrectly grounded entities. We present a neurosymbolic agent that factors long-horizon household tasks into task-directed visual exploration and constrained symbolic planning. In the first phase, a vision-language model and exploration harness acquire goal-relevant predicates and instance bindings from egocentric observations and grounded interactions, producing a symbolic initial state. In the second, a PDDL transition model restricts decoding to tokens that extend applicable actions. Monte Carlo tree search then evaluates executable continuations using a domain-independent planning heuristic. The resulting plans are executable by construction under the transition model, with transfer to the environment conditioned on correct visual grounding. On VirtualHome and ALFWorld, open 4B-27B models exceed 90% success in both environments, and our smallest agent substantially outperforms a 27B direct visual policy in each. Constraints and search prove complementary rather than interchangeable: in ALFWorld either alone solves under a third of tasks, whereas their combination solves over 95%. The method also uses several times fewer generated tokens than extended thinking and far fewer model-visible images than direct interaction, and residual failures localize to state acquisition rather than plan generation without any specialized training.

Problem

• Neurosymbolic Embodied Agents
• Plan Executability
• Visual Grounding
• Environment Dynamics

Innovation

• Neurosymbolic Agent
• Constrained Symbolic Planning
• Monte Carlo Tree Search
• Visual Grounding
• PDDL Transition Model