LACO: Training-Free Latent Communication for Collaborative Driving
Researchers propose LACO, a training-free latent communication paradigm for collaborative driving that addresses high latency and information loss in language-based approaches. The method adapts pretrained driving models to exchange latent representations while avoiding agent identity confusion through a novel fusion strategy.
Key facts
- LACO is a training-free latent communication paradigm for collaborative driving.
- It addresses high latency from autoregressive decoding in language-based communication.
- It prevents information loss from compressing rich representations into discrete tokens.
- Agent identity confusion is identified as a key challenge in direct latent state fusion.
- The method adapts pretrained driving models without additional training.
- The research is published on arXiv with ID 2605.22504.
- Collaborative driving aims to improve safety and efficiency under partial observability.
- The approach uses latent communication instead of language-based reasoning.
Entities
Institutions
- arXiv