Reproduce “CCIL: Context-conditioned imitation learning for urban driving” on OpenResearch

About this paper

Ke Guo, Wei Jing, Junbo Chen, Jia Pan

Researchers from The University of Hong Kong and Alibaba Group developed Context-Conditioned Imitation Learning (CCIL) for urban driving, which predicts future trajectories using only context information and an ego-perturbed goal-oriented coordinate system to counter covariate shift. The method achieved substantially lower collision and off-road rates, and reduced L2 position error compared to various imitation learning baselines on large-scale real-world datasets.

Key insights

• Excluding the ego vehicle's historical state from policy input significantly improves robustness against covariate shift by preventing spurious correlations and leveraging the inherent stability of the broader traffic system.
• The ego-perturbed goal-oriented coordinate system effectively minimizes implicit ego information, which is crucial for reducing policy exploitation of self-motion cues and enhancing long-term driving quality.
• Combining a context-conditioned policy network with an LQR controller ensures both stable, generalizable planning and smooth, comfortable, dynamically feasible trajectories, addressing both safety and comfort aspects.

Summary by alphaXiv.