Chinese Robot Hits 1816 Parcels/Hour in Sorting Test, Beating Figure AI with 70% Lower Hardware Cost

量子位 · wechat · 2026-08-12

In a live, unscripted sorting test without human intervention, a robot developed by Zibian achieved a sorting efficiency of 1,816 parcels per hour. Using a dual-arm setup with standard grippers, it surpassed Figure AI's previous record of 1,848 parcels/hour while reducing hardware costs by 70%.

This performance is powered by their proprietary embodied AI model, WALL-B. Unlike traditional modular VLA architectures, WALL-B uses a Unified World Model (WUM) architecture to natively fuse multi-modal data, enabling the robot to perceive randomly mixed parcels and generate adaptive action strategies in real-time. The robot demonstrated zero-shot generalization across complex items like soft bags and irregular shapes. This validates a highly cost-effective deployment path driven by strong model intelligence rather than complex, expensive hardware, marking a solid breakthrough for Chinese embodied AI in cross-scenario applications.

Original post →

More from Embodied

Embodied channel →