World models test robot software before physical deployment

Baidu’s Apollo World model now acts as an examiner for the company’s driverless cars, simulating complex traffic interactions to validate safety before real-world deployment (Poe Zhao from Hello China Tech, Poe Zhao from Hello China Tech). Baidu architect Yang Ze reports that the model generates camera and lidar data for rare events, allowing the system to complete entire virtual trips from order acceptance to passenger drop-off. This approach treats the virtual city as a high-stakes testing ground where AI agents react to the robotaxi’s maneuvers, providing a standardized environment for performance grading.

X Square Robot’s WALL-SS world model scores candidate robot policies by running them through simulated tasks, effectively ranking software versions before physical testing. In a test of 600 paired runs, the model’s rankings matched physical robot performance with a Pearson correlation of 0.93. While this correlation suggests high predictive power, some researchers argue that real-world deployment remains the only definitive validation, as simulators may fail to capture the full complexity of unstructured physical environments.

These dual applications highlight a shift toward using world models as critical infrastructure for software certification. Whether virtual scores reliably predict success across all edge cases remains the primary point of contention between developers prioritizing rapid iteration and those demanding extensive physical verification.

• NVIDIA’s GR00T 2 model increased success rates on 50 unfamiliar robot tasks from 40% to 95% using five demonstrations per task (🔳 Turing Post).
• Skild and NVIDIA’s Steel solver achieved 3.8 times the simulation throughput of PhysX across 20 tasks, accelerating the testing of robot policies (Alexander Kruel from Axis of Ordinary).