AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
SLoSa: Small, Low-Power, and Safe Foundation Models for Robotic Task Planning
SLoSa addresses safe robotic task planning with foundation-model-based planners that are small, low-power and safety-aware, combining the flexible task interface of LLMs with the formal guarantees of classical planners. The framework builds a safety-aware dataset from classical planners, an advanced teacher LLM and formal verifiers, then applies supervised fine-tuning followed by reinforcement learning fine-tuning guided by a verification-based reward machine. Across four PDDL domains, verifier-guided post-training raises the success rate of compact 4B-14B planners to 70.5-76.0%, and INT4 deployment on a Jetson Orin requires 17.85-26.41 W of average power.