← Back to Robotics
cs.RO

Planning ahead: how LLMs can drive safely despite network delays

Anjie Qiu, Hans D. Schotten

May 21, 2026

Cloud-hosted LLMs provide good driving decisions but are too slow for real-time vehicle control. SteinsGateDrive decouples planning from execution: the LLM generates multiple future scenarios (normal driving, nearby-car interactions, emergency braking) before the control deadline, then a runtime system reuses the best prediction only while safety conditions hold. On highway driving, the system reduced effective latency from 3 seconds to near-zero while maintaining collision-free operation.
Published as Steins;Gate Drive: Semantic Safety Arbitration over Structured Futures for Latency-Decoupled LLM Planning arXiv:2605.22456
Read the original paper →