Research map / Reinforcement learning / Visual and representation RL

CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving

CVPR 2025 · Reinforcement learning › Visual and representation RL · approach: RL for combinatorial routing · 14 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CarPlanner, a consistent auto-regressive RL planner that conditions a PPO-trained trajectory generator on time-invariant longitudinal-lateral mode representations, achieving state-of-the-art closed-loop scores on nuPlan.

Problem

RL-based trajectory planning for autonomous driving suffers from training inefficiency and poor performance on large-scale real-world datasets such as nuPlan, and vanilla auto-regressive planners produce inconsistent multi-modal trajectories.

Approach

Condition an auto-regressive, model-based RL policy on a consistent (time-invariant) longitudinal-lateral mode representation, use an expert-guided displacement-error plus safety reward, an invariant-view module that makes policy inputs time-agnostic, and a rule-augmented selector at inference.

Key results

• On Test14-Random non-reactive, CarPlanner achieves CLS-NR 94.07 vs. 92.07-level PLUTO 91.92 and PDM-Closed 90.05 (improvement of 4.02 over PDM-Closed).
• On Reduced-Val14 non-reactive, CarPlanner achieves CLS-NR 91.45 and S-PR 95.37, surpassing PDM-Closed (91.21 CLS-NR, 92.68 S-PR) and Gen-Drive finetune (87.53).
• Reactive setting CLS-R 91.1 on Test14-Random, slightly below PDM-Closed (91.64); the authors attribute this to training only in non-reactive settings.

Measured effects reported in the paper

Surprising findings

• Ego-history dropout helps IL training but degrades RL performance by harming advantage estimation.
• Backbone sharing between trajectory generator and selector helps IL but hurts RL training due to gradient interference.

Limitations

• RL relies on expert-guided reward design, which may constrain discovery of solutions beyond human expertise.
• Reactive-setting performance slightly below PDM-Closed because the model was trained only in non-reactive settings.
• RL can overfit training environments and degrade in unseen environments.

Experimental setup

Datasets: nuPlan

Benchmarks: nuPlan Test14-Random nuPlan Reduced-Val14

Metrics: CLS-NR CLS-R S-CR S-TTC S-Area S-PR S-Comfort

Baselines: IDM PDM-Closed PDM-Open RasterModel UrbanDriver GC-PGP GameFormer PlanTF PEP PLUTO Gen-Drive

Builds on: PPO PDM PlanTF PLUTO Gen-Drive

Closest papers by meaning

• GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction (ICML 2025 · similarity 0.88)
• AdaWM: Adaptive World Model based Planning for Autonomous Driving (ICLR 2025 · similarity 0.87)
• Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.87)
• Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.87)
• RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.87)