The planning problem constitutes a fundamental aspect of the autonomous driving framework. Recent strides in representation learning have empowered vehicles to comprehend their surrounding environments, thereby facilitating the integration of learning-based planning strategies. Among these approaches, Imitation Learning stands out due to its notable training efficiency. However, traditional Imitation Learning methodologies encounter challenges associated with the co-variate shift phenomenon. We propose Validity Learning on Failures, VL(on failure), as a remedy to address this issue. The essence of our method lies in deploying a pre-trained planner across diverse scenarios. Instances where the planner deviates from its immediate objectives, such as maintaining a safe distance from obstacles or adhering to traffic rules, are flagged as failures. The states corresponding to these failures are compiled into a new dataset, termed the failure dataset. Notably, the absence of expert annotations for this data precludes the applicability of standard imitation learning approaches. To facilitate learning from the closed-loop mistakes, we introduce the VL objective which aims to discern valid trajectories within the current environmental context. Experimental evaluations conducted on both reactive CARLA simulation and non-reactive log-replay simulations reveal substantial enhancements in closed-loop metrics such as Score, Progress, and Success Rate, underscoring the effectiveness of the proposed methodology. Further evaluations against the Bench2Drive benchmark demonstrate that VL(on failure) outperforms the state-of-the-art methods by a large margin.

Traffic congestion is characterized by longer trip times, and increased air pollution. In a static road network, the travel time to a destination is constant and can be computed using the shortest path first algorithm (SPF). However, road network conditions are dynamic, rendering the SPF to perform sub-optimally at times. In addition, in a realistic multiple-vehicle scenario, the SPF routing algorithm can cause congestion by routing all vehicles through the same shortest path. In this paper, we propose a network-aware multi-agent reinforcement learning model for addressing this problem. Our key idea is to assign an RL agent to intersections. Each RL agent operates as a router agent and is responsible for providing routing instructions to approaching vehicles. When a vehicle reaches an intersection, it submits a routing query to the RL agent consisting of its final destination. The RL agent generates a routing response based on (i) the destination, (ii) the current state of the road network, and (iii) routing policies learned by cooperating with other neighboring RL agents. Our experimental evaluation shows that the proposed MARL model outperforms the SPF algorithm by (up to) 20.2% in average travel time.