Bi-Level Optimization Augmented with Conditional Variational Autoencoder for Autonomous Driving in Dense Traffic

Autonomous driving has a natural bi-level structure. The goal of the upper behavioural layer is to provide appropriate lane change, speeding up, and braking decisions to optimize a given driving task. However, this layer can only indirectly influence the driving efficiency through the lower-level trajectory planner, which takes in the behavioural inputs to produce motion commands. Existing sampling-based approaches do not fully exploit the strong coupling between the behavioural and planning layer. On the other hand, end-to-end Reinforcement Learning (RL) can learn a behavioural layer while incorporating feedback from the lower-level planner. However, purely data-driven approaches often fail in safety metrics in unseen environments. This paper presents a novel alternative; a parameterized bi-level optimization that jointly computes the optimal behavioural decisions and the resulting downstream trajectory. Our approach runs in real-time using a custom GPU-accelerated batch optimizer and a Conditional Variational Autoencoder (CVAE) learnt warm-start strategy. Extensive simulations show that our approach outperforms state-of-the-art Model Predictive Control (MPC) and RL approaches in terms of collision rate while being competitive in driving efficiency.

In this work, we propose a bi-level optimization that combines Quadratic Programming (QP) with gradient-free optimization for solving the bi-level problem. Moreover, we train a Conditional Variational Autoencoder (CVAE), embedded with a differentiable optimization layer, for warm-starting our bi-level optimizer. Our approach outperforms state-of-the-art MPC and RL-based baselines in safety metrics in dense traffic.

At the core of our Behavioral Cloning framework lies a combination of feedforward and differentiable optimization layers, conditioned on the observations o, to produce behavioral inputs p that leads to an optimal trajectory as close as possible to expert trajectory demonstrations. CVAE is trained to approximate the complex underlying distribution of optimal trajectories from expert demonstrations that subsumes the behavioral inputs p. The samples drawn from this learned distribution initialize our bi-level optimizer.