opendrivelab/simscale

SimScale is a research project that teaches self-driving car AI systems to drive better by training them on large amounts of simulated driving data alongside real-world driving data. The core problem it addresses is that real-world driving data is expensive and slow to collect, while purely simulated data often does not transfer well to real roads. SimScale bridges that gap with a co-training strategy. The project builds a simulation pipeline that generates diverse, realistic-looking driving scenarios, complete with reactive vehicles that respond to the ego car's actions. These simulated scenes come with pseudo-expert demonstrations, meaning the simulation also provides example trajectories of how a skilled driver would handle each situation. These simulated examples are then mixed with real driving data during training, so the AI model gets the best of both worlds: variety and scale from simulation, real-world texture from actual recordings. The result is an AI driving planner that generalises better to challenging situations it may never have seen in real data alone. The project releases the dataset, pretrained model checkpoints, and training code so researchers can reproduce the results or fine-tune the models on their own planners. It is aimed at autonomous driving researchers who work with end-to-end driving models and want to understand how sim-to-real co-training affects performance at scale.