Imitative Models for Passenger-Scale Autonomous Off-Road Driving

Vision-based control of autonomous vehicles presents major challenges, particularly outside of well struc- tured environments with clear road boundaries and lane mark- ings. Learning control policies from human driving data offers an appealing alternative to classical navigation pipelines: by learning to directly associate observations with actions that avoid obstacles and achieve navigational goals, it is possible to circumvent many of the challenges associated with manually engineering a driving system for unstructured or off-road settings. However, integrating learning-based approaches into robust high-performance control systems presents a major challenge. In this paper, we describe a system for passenger- scale autonomous navigation in off-road environments that combines imitative models with low-level model-predictive con- trol. Although the system learns to control the vehicle directly through perception, it is designed to integrate together learning- based components with constraints and trajectory optimization so as to provide a complete navigational system. Our experi- ments demonstrate the performance of the system in real-world scenarios over complex off-road terrains, and characterize its potential for improvement with the scaling of data collection and interventions. For a video description, see our link here

Recommended citation: Nitish Dashora, Sunggoo Jung, Dhruv Shah, Valentin Ibars, Osher Lerner, Chanyoung Jung, Rohan Thakker, Nicholas Rhinehart, Ali-akbar Agha-mohammadi. (2022). "Imitative Models for Passenger-Scale Autonomous Off-Road Driving." Presented at 2023 International Conference on Intelligent Robots and Systems (IROS).
Download Paper