TinyML-Driven Distributed Collaborative Computing for Autonomous Vehicle Groups in Open Scenes

Existing implementations of autonomous vehicle groups focus on structured environments, overlooking challenges such as centralized computation bottlenecks, environmental interference, and the varied capabilities of vehicles in dynamic open scenes. To address these issues, this paper proposes a TinyML-based distributed framework for autonomous vehicle groups in open scenes. First, a hierarchical collaborative computing architecture is introduced, featuring adaptive inter-vehicle links across four specialized computational units. Second, a method for predicting link quality, which accounts for interference, is developed by using TinyLSTM to assess communication channels dynamically under real-world disturbances. Third, a heterogeneous group formation model is proposed, integrating objectives of computational synergy, stability, and efficiency, and solving it by using an improved discrete particle swarm optimization algorithm. Finally, the simulation results demonstrate that TLPP outperforms existing methods in terms of MAE, FLOPs, and parameters, while DF-A excels in computational synergy, stability, and efficiency.