OpenDriveLab/SimScale

[CVPR 2026 Oral] Learning to Drive via Real-World Simulation at Scale

What it solves

SimScale addresses the challenge of training robust end-to-end autonomous driving planners by providing a scalable way to generate high-fidelity, reactive driving scenarios. It overcomes the limitations of real-world data by synthesizing diverse simulation data and providing a co-training strategy to bridge the gap between simulated and real-world environments.

How it works

The project implements a scalable simulation pipeline that creates synthetic driving scenarios featuring "pseudo-expert" demonstrations. It utilizes a sim-real co-training strategy to improve the robustness and generalization of end-to-end planners. This co-training can be performed using either pseudo-expert demonstrations or by relying on rewards only. The system is designed to work with various planner backbones, such as ResNet34, and is compatible with datasets like nuPlan and NAVSIM.

Who it’s for

It is designed for researchers and engineers working on end-to-end autonomous driving, specifically those focusing on motion planning and the scaling properties of sim-real learning systems.

Highlights

  • Scalable Simulation Pipeline: Synthesizes high-fidelity reactive driving scenarios with pseudo-expert demonstrations.
  • Sim-Real Co-training: A strategy that synergistically improves robustness and generalization across different end-to-end planners.
  • Comprehensive Scaling Insights: Provides a recipe for understanding the scaling properties of sim-real learning for autonomy.
  • Diverse Model Support: Includes pre-trained checkpoints for various agents like LTF, DiffusionDrive, and GTRS-Dense.

Related

  • Project
    EzioBy/Ditto

    Ditto is a framework for generating high-quality synthetic video editing data to train Editto, a state-of-the-art instruction-based video editing model.

  • Project
    swc-17/SparseDrive

    SparseDrive is an end-to-end autonomous driving framework that uses sparse scene representation to unify perception and planning, significantly improving inference speed and reducing collision rates.

  • Project
    NVIDIA-NeMo/Skills

    A collection of pipelines for LLM development that streamlines synthetic data generation, training, and evaluation across local workstations and large-scale GPU clusters.

  • Project
    liruiw/GenSim

    GenSim is an LLM-powered pipeline that automatically generates robotic simulation tasks and environments to augment training data for multitask robotic policies.

  • Project
    Lei-Kun/RL-100

    A real-world reinforcement learning framework for robotic manipulation that enables a unified pipeline from imitation learning to offline/online RL and one-step policy distillation.