NAVSIM v2: Pseudo-Simulation for Autonomous Driving & ICCV 2025 Challenge Winner Presentation

Event: CVPR 2025 Workshop on Learning to See: Advancing Spatial Understanding for Embodied Intelligence · Duration: 0 min · ▶ Watch on YouTube

Abstract

This workshop session introduces NAVSIM v2, a novel pseudo-simulation framework designed for efficient and comprehensive evaluation of autonomous driving systems. It combines real-world observations with pre-generated synthetic data to enable parallel model inferencing while capturing complex phenomena like compounding errors and causal confusion, typically only found in computationally intensive closed-loop simulations. The session also presents the results of the ICCV 2025 Autonomous Grand Challenge, highlighting the winning solution, GTRS (Generalized Trajectory Scoring), developed by NVIDIA. GTRS leverages pseudo-reinforcement learning, diffusion-policy-generated trajectories, and advanced scoring techniques to achieve high performance in end-to-end multimodal planning.

Speakers

  • Wei Cao — NAVSIM Team
  • Zhenxin Li — NVIDIA, Fudan University

Talks (2)

  • 00:00:00 — Wei Cao: NAVSIM v2: Pseudo-Simulation for Autonomous Driving
    • Introduces NAVSIM v2 as a pseudo-simulation framework for autonomous driving evaluation, bridging the gap between compute-efficient open-loop and comprehensive but intensive closed-loop evaluations, and presents the ICCV 2025 Challenge results.
  • 09:27:00 — Zhenxin Li: GTRS: Generalized Trajectory Scoring for End-to-end Multimodal Planning
    • Presents GTRS, the winning solution for the Autonomous Grand Challenge, leveraging pseudo-reinforcement learning, diffusion-policy-generated trajectories, and advanced scoring techniques for high-performance end-to-end multimodal planning.

Key Takeaways

  • Pseudo-simulation offers a compute-efficient alternative to traditional closed-loop simulations for autonomous driving evaluation, addressing limitations of open-loop methods while modeling complex real-world scenarios.
  • The EPDMS metric effectively combines safety, compliance, and comfort aspects for comprehensive trajectory evaluation, showing strong correlation with actual closed-loop scores.
  • The winning GTRS solution demonstrates the power of combining diffusion models for diverse trajectory generation with robust, generalized scoring mechanisms and ensemble techniques.
  • Enhancing trajectory vocabulary with dynamic, scene-aware trajectories and employing regularization techniques like dropout and sensor augmentation significantly improves model performance and generalization in autonomous driving planning.

Methods / Models / Datasets Mentioned

  • NAVSIM v2
  • Extended Predictive Driver Model Score (EPDMS)
  • 3D Gaussian Splatting (3DGS)
  • GTRS (Generalized Trajectory Scoring)
  • HydraMDP
  • Diffusion Policy Model
  • GTRS-Dense
  • GTRS-Aug

Topics

Autonomous Driving · Pseudo-Simulation · Open-Loop Evaluation · Closed-Loop Evaluation · Trajectory Planning · Reinforcement Learning · Diffusion Models · Sensor Augmentation · Generalization · End-to-End Driving · Challenge

Notes

Open for commentary — connections to other work, critiques, follow-up reading.