Rethinking Imitation-based Planner for Autonomous Driving

TL;DR

The paper addresses the lack of standardized benchmarks for imitation-based driving planners by leveraging nuPlan to study ego feature design and data augmentation while revealing a notable imitation gap. It demonstrates that history-based inputs can cause shortcut learning and introduces an attention-based state dropout encoder to mitigate this, achieving strong closed-loop performance. By examining augmentation strategies and introducing an RL Adapter to bridge dynamics, the work delivers PlanTF, a pure imitation-based baseline that rivals rule-based and post-optimized methods on nuPlan benchmarks. The work provides open benchmarks and models to facilitate fair comparisons and future advancements in imitation-learning for autonomous driving.

Abstract

In recent years, imitation-based driving planners have reported considerable success. However, due to the absence of a standardized benchmark, the effectiveness of various designs remains unclear. The newly released nuPlan addresses this issue by offering a large-scale real-world dataset and a standardized closed-loop benchmark for equitable comparisons. Utilizing this platform, we conduct a comprehensive study on two fundamental yet underexplored aspects of imitation-based planners: the essential features for ego planning and the effective data augmentation techniques to reduce compounding errors. Furthermore, we highlight an imitation gap that has been overlooked by current learning systems. Finally, integrating our findings, we propose a strong baseline model-PlanTF. Our results demonstrate that a well-designed, purely imitation-based planner can achieve highly competitive performance compared to state-of-the-art methods involving hand-crafted rules and exhibit superior generalization capabilities in long-tail cases. Our models and benchmarks are publicly available. Project website https://jchengai.github.io/planTF.

Figures (5)

• Figure 1: A brief overview of our baseline model. Agents, map, and ego features are separately encoded and then concatenated, which are subsequently processed by a stack of transformer encoder layers. The baseline model jointly predicts traffic agents and plans for ego vehicle at the scene level.
• Figure 2: The left side shows the planning trajectory of the state6 model by adjusting AV's steering angle from 0.15 to 0.5 rad. The right side illustrates the magnitude of the gradient concerning the trajectory endpoint's position in relation to the AV's kinematic states.
• Figure 3: Illustration of the attention-based state dropout encoder.
• Figure 4: (a) The original scenario. (b) Random noise is added to the AV's current state and history motion is smoothed. (c) The coordinates of the scenario are re-normalized based on the perturbed position of the AV. (d) A corrected future trajectory is generated using constrained nonlinear optimization.
• Figure 5: Illustration of the imitation gap and the proposed RL adapter.