Overview
- Apollo World, which Baidu unveiled on Tuesday, is an agent-driven multimodal world model that claims to produce up to 100,000 km of simulated driving per day and stream camera and lidar data in real time.
- The model is trained on driving history from Apollo Go and other sources totaling more than 350 million kilometers and is already being used in Apollo Go’s training pipeline for planned international deployments.
- Baidu proposed a new '2+3' evaluation standard for judging world models that is meant to set a common way to measure simulation fidelity and readiness for physical systems.
- Industry experts warn that large-scale simulation still requires heavy capital for model training and vehicle deployment, and rivals such as XPeng, WeRide, Pony.ai and Tesla are racing to commercialize robotaxis.
- World models generate rare and dangerous scenarios at scale to reduce on-road testing but regulators, safety verification and proof that simulated learning transfers to real-world driving remain the key hurdles to wider rollouts.