Neurosymbolic Embodied Agents
The paper reports over 90% task success by forcing embodied plans through symbolic constraints and search.
The agent first builds a symbolic state from egocentric visual exploration, then uses a PDDL transition model to restrict action decoding. Monte Carlo tree search scores executable continuations with a planning heuristic. On VirtualHome and ALFWorld, open 4B-27B models exceed 90% success, and the smallest setup beats a 27B direct visual policy. The authors say remaining failures mostly come from state acquisition, not plan generation. ArXiv · AI/CL/LG's note
The agent first builds a symbolic state from egocentric visual exploration, then uses a PDDL transition model to restrict action decoding. Monte Carlo tree search scores executable continuations with a planning heuristic. On VirtualHome and ALFWorld, open 4B-27B models exceed 90% success, and the smallest setup beats a 27B direct visual policy. The authors say remaining failures mostly come from state acquisition, not plan generation. ArXiv · AI/CL/LG's note
score 4