Overview
- Baidu launched Apollo World on Tuesday as an agent-driven, multimodal world model designed to simulate traffic and predict road behavior to train autonomous vehicles.
- The system can produce up to 100,000 km of simulated driving each day and streams real-time camera and lidar data so self-driving stacks can sense and navigate virtual environments.
- Apollo World is built on more than 350 million km of Apollo Go real-world driving data and is already integrated into Apollo Go to train robotaxis for international operations.
- Baidu says the platform will cut reliance on slow and costly live testing, but industry experts note companies still face heavy capital costs for training models and deploying vehicles.
- Baidu proposed a new '2+3' evaluation standard for world models and the company's push could shape benchmarking as rivals such as XPeng plan robotaxi rollouts and analysts forecast large fleet growth by 2035.