End-to-end Driving via Conditional Imitation Learning (1710.02410v2)

Abstract: Deep networks trained on demonstrations of human driving have learned to follow roads and avoid obstacles. However, driving policies trained via imitation learning cannot be controlled at test time. A vehicle trained end-to-end to imitate an expert cannot be guided to take a specific turn at an upcoming intersection. This limits the utility of such systems. We propose to condition imitation learning on high-level command input. At test time, the learned driving policy functions as a chauffeur that handles sensorimotor coordination but continues to respond to navigational commands. We evaluate different architectures for conditional imitation learning in vision-based driving. We conduct experiments in realistic three-dimensional simulations of urban driving and on a 1/5 scale robotic truck that is trained to drive in a residential area. Both systems drive based on visual input yet remain responsive to high-level navigational commands. The supplementary video can be viewed at https://youtu.be/cFtnflNe5fM

• The paper demonstrates that high-level command integration in imitation learning significantly improves autonomous driving performance.
• It introduces two architectures—command input and command branching—with the latter excelling in complex urban scenarios.
• The study validates the approach through simulations and scaled physical tests, emphasizing the impact of noise injection and data augmentation.

End-to-end Driving via Conditional Imitation Learning

This essay provides an academic overview of the paper titled "End-to-end Driving via Conditional Imitation Learning" by Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. The paper addresses a key limitation in imitation learning applied to autonomous driving by introducing a conditional component that provides high-level commands to guide vehicle actions, thereby enhancing both control and performance in complex urban environments.

Introduction and Motivation

Imitation learning in autonomous driving typically involves training models on human driving demonstrations with the aim of replicating expert behaviors. Such models, once trained, should map perceptual inputs (e.g., images from a forward-facing camera) directly to control commands (e.g., steering, acceleration). While successful in tasks like lane following and obstacle avoidance, traditional imitation learning approaches lack the capability to follow specific navigational commands at intersections. This oversight restricts the broader applicability and usability of these models, as they cannot be directed to perform tasks such as making specific turns as dictated by human operators or high-level planners such as GPS navigation systems.

Conditional Imitation Learning

The solution proposed by the authors is conditional imitation learning, where the learned driving policy is conditioned on high-level navigational commands provided during both training and testing phases. This enhanced formulation enables the model to not only imitate low-level driving actions but also respect high-level directives like "turn left at the next intersection."

Methodology and Network Architecture

The core contribution of the paper lies in the network architecture designed to integrate conditional inputs. Two architectures were explored:

1. Command Input Architecture (CIA): This approach takes the high-level command as an additional input along with the visual and measurement data. The command is processed alongside these inputs and integrated into the control decision framework.
2. Command Branching Architecture (CBA): This architecture uses the command to select among specialized branches within the network, each tailored to different navigational commands (e.g., going straight, turning left, turning right).

The authors found that the branching architecture (CBA) generally performed better, providing more robust responses to high-level commands and handling complex urban driving tasks with greater reliability.

Experimental Evaluation

Simulation Studies

The effectiveness of conditional imitation learning was evaluated using the CARLA simulator, a high-fidelity simulation environment for urban driving. The paper included two towns, one used for training and the other for testing, to robustly assess the model's generalization capabilities. The results indicated that the command-conditional methods significantly outperformed non-conditional counterparts in terms of success rates for driving tasks and the distance traveled without infractions.

Physical System Validation

The approach was further validated on a 1/5 scale robotic truck equipped with a visual sensor suite, embedded computing hardware, and a control system. In real-world tests conducted in a residential area, the command-conditional methods achieved notable success in following high-level directives. Crucially, the model trained using noise injection and data augmentation exhibited higher robustness and generalization capabilities, performing well in varying environmental conditions.

Key Findings and Implications

The paper reports several critical findings:

• Command-conditional models significantly outperformed standard and goal-conditional models in terms of task success rates and robustness against disturbances.
• Training architectures that incorporate explicit high-level commands enable the model to resolve ambiguities in perceptual inputs leading to more reliable autonomous driving.
• Data augmentation and noise injection during training are crucial for robust real-world performance.

These findings have profound implications for both theoretical research and practical deployments in autonomous driving. The ability to condition imitation learning on high-level commands can bridge a critical gap, making autonomous systems more interactive and controllable by human operators or high-level automated planners.

Future Directions

Future research can explore several directions stemming from this work:

1. Increasing the complexity and scalability of the network architectures to handle broader and more varied urban scenarios.
2. Integrating natural language processing capabilities to allow high-level commands to be given in natural language, enhancing user interaction.
3. Extending the conditional imitation framework to other robotic domains where high-level guidance is crucial for task performance.

Conclusion

The paper "End-to-end Driving via Conditional Imitation Learning" presents a significant methodological advancement in the domain of autonomous driving by incorporating high-level commands into the imitation learning framework. This innovation facilitates the creation of more flexible and responsive driving policies, capable of performing complex maneuvers in dynamic environments. The approach shows promise for scaling up to full-fledged autonomous urban driving systems, offering both practical benefits for real-world applications and a fertile ground for future academic research.