Xinghaitu Showcases G0.5 Model at WRC, Addressing Robot Generalization and Precision

量子位 · wechat · 2026-08-20

At the World Robot Conference (WRC), Xinghaitu demonstrated its progress in 'Productivity Awakening' for embodied intelligence across a 500-square-meter booth, focusing on the leap from robots 'being able to work autonomously' to 'working fast, accurately, and with generalization'.

Technical Core: G0.5 VLA Model

Unified Autoregressive Architecture: Integrates visual understanding, embodied reasoning, and action generation into a single Transformer Decoder. This enables continuous generation of 'thinking' and 'doing', supporting long-horizon tasks (e.g., end-to-end picking and packing in JD front warehouses).
Generalization Capability: For both flexible objects (like folding boxes, folding clothes) and rigid objects in the real world, G0.5 can adjust actions based on real-time status and has generalized to over 10,000 SKUs.

Three Pillars of Productivity:

Generalization: Adapts to unseen objects and terrains (e.g., bipedal robot Kengo) without relying on specific environments.
Precision: In industrial assembly scenarios, combined with reinforcement learning, operation precision is controlled within 1mm with a success rate of 99.9%, boosting efficiency by 400%.
Efficiency: In real business flows (e.g., front warehouses), end-to-end task completion time is kept within an acceptable range.

Data and Flywheel: Xinghaitu insists on a full-stack self-research approach (hardware, model, data loop). By collecting high-quality real-world data through proprietary robots (R1, Kengo, etc.), it drives the collaborative iteration of models and hardware.

Original post →

Xinghaitu Showcases G0.5 Model at WRC, Addressing Robot Generalization and Precision

More from Companies & People