$AutoDrive\text{-}P^3$: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning

Dimensions overall score 8.0

• AutoDrive-P3 achieves state-of-the-art planning performance on both open-loop (nuScenes) and closed-loop (NAVSIMv1/v2) benchmarks.result 90%
  Evidencepartial

  Extensive experiments on both open-loop (nuScenes) and closed-loop (NAVSIMv1/v2) benchmarks demonstrate that our approach achieves state-of-the-art performance in planning tasks.

  Implicationpartial

  Explicitly stated in the abstract and supported by quantitative results in tables comparing against other methods.

  Verificationpartial

  partial

• The proposed P3-GRPO hierarchical reinforcement learning algorithm significantly enhances planning performance by providing progressive supervision across perception, prediction, and planning tasks.method 85%
  Evidencepartial

  More importantly, experimental results validate that our proposed P3-GRPO algorithm significantly enhances planning performance through its hierarchical and progressive supervision mechanism, which systematically improves perception and prediction capabilities and consequently leads to more reliable and

  Implicationpartial

  Directly stated in the analysis section with a description of its mechanism, though specific performance improvement numbers are not provided in the excerpt.

  Verificationpartial

  partial

• Current VLM-based autonomous driving approaches suffer from either bypassing perception/prediction stages or using fragmented, non-synergistic modules, compromising decision-making.limitation 95%
  Evidencepartial

  However, current VLM-based approaches suffer from two major limitations: 1) Some VLMs directly output planning results without chain-of-thought (CoT) reasoning, bypassing crucial perception and prediction stages which creates a significant domain gap and compromises decision-making capability; 2) Other VLMs can generate outputs for perception, prediction, and planning tasks but employ a fragmented decision-making approach where these modules operate separately, leading to a significant lack of synergy that undermines true planning performance.

  Implicationpartial

  Explicitly and clearly stated as a core limitation in the abstract, forming the primary motivation for the work.

  Verificationpartial

  partial

• AutoDrive-P3 introduces a unified chain-of-thought (P3-CoT) that seamlessly integrates Perception, Prediction, and Planning through structured reasoning.method 90%
  Evidencepartial

  To address these limitations, we propose AutoDrive-P3, a novel framework that seamlessly integrates Perception, Prediction, and Planning through structured reasoning.

  Implicationpartial

  Directly stated as the core contribution in the abstract and methodology section.

  Verificationpartial

  partial

• The framework introduces dual thinking modes (detailed and fast) to balance inference efficiency with performance.technical 85%
  Evidencepartial

  Additionally, to balance inference efficiency with performance, we introduce dual thinking modes: detailed thinking and fast thinking.

  Implicationpartial

  Explicitly mentioned in the abstract as a feature, and results are provided for both modes.

  Verificationpartial

  partial

• The model is initialized via a cold-start supervised fine-tuning (SFT) phase using the P3-CoT dataset to learn the necessary CoT answer format for autonomous driving.method 80%
  Evidencepartial

  We first cold start the base model using P3-CoT to make up for the gap between VLM and autonomous driving and learn the CoT answer format.

  Implicationpartial

  Described in the methodology section and Figure 4 caption, though the specific details of the dataset are not in the excerpt.

  Verificationpartial

  partial

• AutoDrive-P3 outperforms existing VLM-based methods like DriveVLM, OmniDrive, and OpenDriveVLA on the nuScenes benchmark metrics (e.g., minADE, minFDE).result 95%
  Evidencepartial

  AutoDrive-P3 (Ours-Detailed) 0.15 0.30 0.54 0.33 0.00 0.02 0.15 0.06 Qwen2.5-VL-3B

  Implicationpartial

  Supported by direct numerical comparisons in the provided results table, showing lower error metrics.

  Verificationpartial

  partial

Competitive landscape

A unified framework for autonomous driving that integrates perception, prediction, and planning through chain-of-thought reasoning and reinforcement learning, achieving state-of-the-art performance.