Event
SLoSa framework trains small, low-power safety-aware robot planning models
1 reports · 1 sources
Weizhe Xu et al. propose SLoSa, a neuro-symbolic compression and fine-tuning framework that uses classical planners, teacher large language models, and a formal verifier to build a safety-aware dataset, then applies supervised fine-tuning and reinforcement learning fine-tuning based on a verification reward machine to train small, low-power, safety-aware robot task planning models. In evaluations across four PDDL planning domains, verifier-guided post-training raised the success rate of small planners with 4B-14B parameters to 70.5-76.0%; INT4 quantized deployment on Jetson Orin had average power consumption of 17.85-26.41 W. The evaluation also covered two household simulation environments, a local edge platform, and a desktop robotic arm.
Report timeline
- SLoSa: A Small, Low-Power, Safety-Aware Foundation Model Framework for Robotic Task Planning
SLoSa is a neuro-symbolic compression and fine-tuning framework that builds a safety-aware dataset using classical planners, a teacher large language model, and formal verifiers, then applies supervised fine-tuning and reinforcement learning fine-tuning guided by a verification-based reward machine to train small, low-power, safety-aware robotic task planning models.
arXiv cs.RO ·