Three-month-old startup Lingxi Zhiyong takes national third in industrial humanoid contest with ROSS Harness

The second World Humanoid Robot Games divided its 51 events into two evaluation tracks: competition events test physical limits and coordination, while scenario events are benchmarked directly against real job standards. The industrial assembly and loading station is the closest to an actual production line, judged on long-horizon stability, millimeter-level precision and autonomous recovery under disturbance.

In that factory-adjacent event, Lingxi Zhiyong — founded just three months ago — moved through the preliminaries and finals with a robot assembled from a demo-grade chassis, finishing in the national top three with a score of 160 and becoming the only robotics company besides established industry leaders to win in the industrial scenario category.

The result is notable because the company is young, its entry was not a mass-production model but a demo-grade assembly, and its competitors were established brands with deep experience in chassis and robotic arms. Winning without hardware advantages, Lingxi Zhiyong relied not on the body but on its self-developed ROSS Harness Agent.

The win sits against a broader industry pain point: no matter how polished a leader's demo is, real industrial deployments still fail delivery checks, stuck on unstable success rates, insufficient throughput, expensive operations and difficult migration. The root cause, the company argues, is the methodological limit of a single-model approach — models are probabilistic outputs with a ceiling on single-shot success, while industrial scenes are open worlds where existing systems cannot self-evolve and experience cannot be沉淀 into reusable assets.

The self-developed ROSS Harness, led by co-founder and CTO Dr. Duan Yifan and named after cybernetics founder W. Ross Ashby, is an industrial-grade execution system that builds a continuous closed loop between model and robot, converting probabilistic model actions into stable, controllable, reusable production capability. Five core capabilities — model abstraction, skill abstraction, agentic AI, layered safety monitoring and a data flywheel — together form the underlying engine of industrial-grade self-evolution. "Self-evolution" here does not mean retraining a large model each time; instead, models, skills, tools, task scheduling and memory co-evolve in real industrial tasks.

Lingxi Zhiyong is betting its moat on the harness layer rather than hardware scale, but it has not abandoned model work: its self-developed CONWAY industrial-native model, focused on local action decisions in industrial scenes, has produced results in force sensing, fast inference, action alignment and safety control that feed the harness's design and evolution. The team's core judgment is that model capability and harness are not substitutes — the stronger the foundation model, the more the harness can convert its abilities into stable, controllable, scalable industrial execution; the model decides a robot's initial capability, while the harness decides whether that capability keeps growing once it enters the production line.

The team is a mentor-student duo from USTC: Duan Yifan, a USTC PhD who focused on embodied and spatial intelligence during his doctorate, has published 36 papers with over 390 citations at venues like RSS, IJRR, TRO, ICRA and IROS, with full-stack engineering experience from autonomous driving to industrial embodiment; chief scientist Professor Ji Jianmin spent more than a decade leading USTC's Kejia and Jiajia service-robot cognitive decision systems, won the RoboCup@Home world championship, and previously served as CTO of Nullmax and chief scientist at Pangolin Robotics. The team has published more than 100 embodied-AI papers at top venues with over 1,000 total citations, and has validated deployments in machining load/unload, precision screw driving and stator pressing.

Next, Lingxi Zhiyong is pushing the self-evolution engine toward productization — delivering competition-proven harness capabilities to more industrial scenarios through a standardized hardware platform, so every robot entering a production line can keep iterating skills, optimizing strategies and self-repairing, truly becoming "stronger with use." In its view, after the emergence of large language models, the most phenomenal products were not chatbots but agents like Codex and Claude Code — the same foundation model performs very differently inside different harnesses, and the physical AI world is replaying that logic: the next golden track in industrial embodied intelligence belongs to players who evolve "Model + Harness" together.