[Paper Review] ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst

Mayank Bansal, Alex Krizhevsky+1 more|arXiv (Cornell University)|Dec 7, 2018

Autonomous Vehicle Technology and SafetyEngineering106 citations

TL;DR

The paper presents ChauffeurNet, a mid-level imitation learning approach for autonomous driving that augments expert data with perturbed trajectories and environment-focused losses to achieve closed-loop robustness, including real-world deployment.

ABSTRACT

Our goal is to train a policy for autonomous driving via imitation learning that is robust enough to drive a real vehicle. We find that standard behavior cloning is insufficient for handling complex driving scenarios, even when we leverage a perception system for preprocessing the input and a controller for executing the output on the car: 30 million examples are still not enough. We propose exposing the learner to synthesized data in the form of perturbations to the expert's driving, which creates interesting situations such as collisions and/or going off the road. Rather than purely imitating all data, we augment the imitation loss with additional losses that penalize undesirable events and encourage progress -- the perturbations then provide an important signal for these losses and lead to robustness of the learned model. We show that the ChauffeurNet model can handle complex situations in simulation, and present ablation experiments that emphasize the importance of each of our proposed changes and show that the model is responding to the appropriate causal factors. Finally, we demonstrate the model driving a car in the real world.

Motivation & Objective

• Demonstrate that pure behavior cloning with mid-level inputs/outputs is insufficient for robust autonomous driving in complex scenarios.
• Introduce data augmentation via synthesized trajectory perturbations to expose the model to non-expert behaviors (e.g., collisions, off-road) and improve safety.
• Propose a multi-loss training framework that combines imitation losses with environment losses to guide robust behavior in closed-loop driving.
• Show that a mid-level representation pipeline can train a driving policy that operates in simulation and transfers to real-world vehicle tests.

Proposed method

• Use mid-level, top-down representations (roadmap, traffic lights, speed limit, route, dynamic objects) as input to ChauffeurNet.
• Employ a Memory-enabled AgentRNN that iteratively predicts future waypoints and a per-pixel agent-bounding-box heatmap.
• Co-train Road Mask Network and PerceptionRNN to predict drivable areas and future positions of other agents using shared features.
• Train with a combination of imitation losses (position, heading, box, sub-pixel refinements, speed) and environment losses (collision, on-road, geometry, object predictions, road mask).
• Introduce Past Motion Dropout to prevent cheating on past history, and Trajectory Perturbation to generate challenging, perturbed training examples.
• Apply Imitation Dropout (randomly zeroing imitation loss) to improve robustness over plain loss weighting.

Experimental results

Research questions

• RQ1 Can pure imitation learning with mid-level representations reach robust driving performance in closed-loop settings?
• RQ2 Does augmenting imitation with perturbed trajectories and environment-focused losses improve safety and robustness in complex driving scenarios?
• RQ3 How do past-motion dropout, trajectory perturbations, and imitation dropout affect learning and generalization?
• RQ4 Can a mid-level perception/control decomposition transfer from simulation/training data to real-world vehicle operation?

Key findings

• Pure imitation learning with 30 million logged examples is insufficient for robustness in closed-loop driving.
• Perturbing expert trajectories and including environment losses significantly improves handling of challenging scenarios (e.g., nudging around parked cars, recovering from lane departures).
• Imitation dropout outperforms simple loss re-weighting, yielding better safety and robustness in ablation tests.
• A model trained with perturbations and environment losses can drive in the real world and negotiate turns, stops, and traffic signals.
• The system operates with mid-level representations and a trajectory-based controller, enabling closed-loop testing in simulation before real deployment.