SwarmDrive: Semantic V2V Coordination for Autonomous Driving
SwarmDrive is a semantic Vehicle-to-Vehicle (V2V) coordination framework that uses local Small Language Models (SLMs) to reduce latency in autonomous driving. Instead of relying on cloud-hosted LLM inference, which introduces round-trip delays and requires stable connectivity, SwarmDrive enables nearby vehicles to share compact intent distributions only when uncertainty is high, fusing them through event-triggered consensus. In a 5-seed executable study around an occluded intersection case, SwarmDrive under 6G communication raised success from 68.9% to 94.1% over a single local SLM, reducing latency from 510 ms (cloud) to 151.4 ms. However, increased vehicle participation leads to higher communication overhead and packet loss. The framework was evaluated using matched operating-point comparisons and robustness sweeps.
Key facts
- SwarmDrive uses local Small Language Models (SLMs) for V2V coordination.
- Cloud-hosted LLM inference adds round-trip delay and requires stable connectivity.
- Vehicles share compact intent distributions only when uncertainty is high.
- Event-triggered consensus fuses shared intent distributions.
- Study used a 5-seed executable setup with one occluded intersection case.
- SwarmDrive under 6G raised success from 68.9% to 94.1% over a single local SLM.
- Latency reduced from 510 ms (cloud) to 151.4 ms.
- More participating vehicles increase communication overhead and packet loss.
Entities
—