Learning from All Vehicles (2203.11934v3)

Abstract: In this paper, we present a system to train driving policies from experiences collected not just from the ego-vehicle, but all vehicles that it observes. This system uses the behaviors of other agents to create more diverse driving scenarios without collecting additional data. The main difficulty in learning from other vehicles is that there is no sensor information. We use a set of supervisory tasks to learn an intermediate representation that is invariant to the viewpoint of the controlling vehicle. This not only provides a richer signal at training time but also allows more complex reasoning during inference. Learning how all vehicles drive helps predict their behavior at test time and can avoid collisions. We evaluate this system in closed-loop driving simulations. Our system outperforms all prior methods on the public CARLA Leaderboard by a wide margin, improving driving score by 25 and route completion rate by 24 points. Our method won the 2021 CARLA Autonomous Driving challenge. Code and data are available at https://github.com/dotchen/LAV.

• The paper presents an innovative end-to-end driving system, LAV, that integrates observations from all vehicles to enrich training signals.
• LAV employs an RNN-based motion planner and PID-controlled vehicle-aware control to improve trajectory prediction and reduce collisions.
• Experiments in CARLA demonstrate that using multi-vehicle data boosts driving scores by 25 points and route completion by 24 points.

Learning from All Vehicles: An Overview

This paper presents a novel approach to train driving policies by leveraging observations from all vehicles in the environment, rather than solely focusing on data gathered from the ego-vehicle. The essence of this work is the ability to construct diverse driving scenarios through the behaviors of other agents without the need for additional data collection. This is achieved by developing a framework known as Learning from All Vehicles (LAV), which enhances the richness of training signals by providing an intermediate representation that remains unaffected by the viewpoint of the controlling vehicle.

Methodology

The LAV system is an end-to-end, mapless driving framework divided into three integral components:

1. Perception Module: Utilizing multi-modal sensors, this module constructs a representation invariant to the observer's perspective. It employs auxiliary tasks such as 3D detection and segmentation, ensuring the feature representation is consistent across different vehicles. This invariant representation is crucial for both perception and subsequent motion planning.
2. Motion Planning: A key contribution of this work is the formulation of a motion planner that can train on trajectories of both the ego-vehicle and surrounding vehicles. The motion planner uses an RNN-based approach to predict future waypoints, while leveraging supervised signals from all visible vehicles to improve prediction accuracy. A privileged distillation method ensures the planner learns from ground-truth perception outputs, refining its trajectory predictions over iterative cycles.
3. Vehicle-aware Control: To convert motion plans into actionable steering and acceleration commands, a PID controller is employed. It uses predictions not only from the ego-vehicle but also incorporates potential collision scenarios predicted for surrounding vehicles.

Experiments and Results

The system is validated using the CARLA simulation environment, where LAV achieved leading results on the public CARLA Leaderboard. Specifically, it improved the driving score by 25 points and route completion rate by 24 points compared to existing methods. The rigorous evaluation demonstrated that the approach effectively reduces errors related to collisions and enhances route navigation under diverse conditions.

Ablation Studies

The paper provides comprehensive ablation studies assessing the impact of core components:

• Inclusion of other vehicles' data significantly improves performance metrics, highlighting the efficacy of training with a broader data set.
• Privileged distillation proves instrumental in reducing training noise and improving model robustness.
• The method's efficacy diminishes slightly when extending to vehicles beyond a 15-meter range, suggesting limitations in long-distance data effectiveness.

Implications and Future Directions

From a practical perspective, the LAV framework holds significant potential for real-world autonomous driving applications by enhancing the adaptability of driving policies to diverse scenarios. Theoretically, it provides insights into handling partial observability and multi-agent learning within the field of AI. Future developments could focus on probabilistically modeling diverse actor behaviors and exploring beyond raster-based perception, potentially incorporating such systems into real-world settings after thorough testing in varied and complex scenarios.

This paper contributes to advancing the domain of autonomous driving by presenting a robust and adaptable framework capable of learning from the collective driving experiences of all vehicles, paving the way for more comprehensive and intelligent driving solutions.