NSDL: Neuro-Symbolic Division of Labor for Multi-Robot Task Planning

Planning for a heterogeneous robot team from a scene image and an instruction involves two kinds of decisions. Some must be judged from the scene: which goals are already met, how the task splits, and which robot does what. Others follow from these and can be computed: prerequisite actions, handoffs between robots, and timing.

In NSDL, a small vision-language model learns the first kind and writes its judgments as a plan sketch, and a deterministic solver expands the sketch into an executable joint plan. The model calls two tools, a detector for evidence and the solver, and is trained with the solver in the loop. We evaluate on VIKI-L2 and on LOCUS-Bench, a new benchmark that grades the two kinds of difficulty separately. The 7B model reaches 97.0% out-of-distribution accuracy on VIKI-L2.

The paper is on arXiv. Code is available on GitHub; model weights and LOCUS-Bench are on Hugging Face.

Recommended citation: Xincheng Tan, Ke Wang, Yining Di, Minfeng Zhu^†. NSDL: Neuro-Symbolic Division of Labor for Multi-Robot Task Planning. arXiv preprint arXiv:2611.16671, 2026.
Download Paper