Interesting Engineering

China’s auto giant unveils 1,400 TOPS robotaxi built on human-like reasoning

Back to overview

Chinese automotive giant Geely has introduced its first purpose-built robotaxi prototype, the Eva Cab. Unlike retrofitted vehicles, the model is designed from the ground up around artificial intelligence, integrating a 196-billion-parameter Step 3.5 model with the company’s H9 high-level autonomous driving system. The platform delivers up to 1,400 TOPS of computing power, enabling real-time data processing and rapid decision-making. According to Geely, the system reaches inference speeds of 350 TPS and can react up to three times faster than a human driver. The company claims the Eva Cab is capable of handling 99% of everyday driving scenarios , including more unpredictable environments like manual toll booths and unmarked rural roads, where it can effectively anticipate viable routes. World action model brings human-like reasoning to Geely’s robotaxi Driving the system is Geely’s World Action Model (WAM), which reimagines the conventional perception-to-decision pipeline as a continuous closed-loop architecture. By layering macro-level route planning with micro-level real-time reasoning, the platform enables the vehicle to interpret and respond to dynamic environments with greater nuance. The result is behavior that more closely mirrors that of an experienced human driver, particularly in ambiguous situations such as unstructured roads or complex traffic interactions where negotiation and anticipation are critical, CarNewsChina reports . Geely has packed the prototype with an extensive sensor suite, built around 43 perception components that include LiDAR and high-definition cameras. Together, they create a triple-layered 360-degree field of view designed to eliminate blind spots and maintain constant awareness of the vehicle’s surroundings. The system continuously maps nearby activity, such as detecting pedestrians, vehicles, and unexpected obstacles. In internal testing, Geely says the platform performs reliably in demanding urban scenarios, including complex mul…