Startup Lingxi Zhiyong takes third place at humanoid robot games with model-plus-Harness approach
机器之心 · wechat · 2026-08-27
At the 2026 World Humanoid Robot Games' industrial assembly event, Lingxi Zhiyong, a Shanghai startup founded by a USTC PhD born after 1998, scored 160 points and ranked third among all companies in its debut—behind only two teams from leading robotics firms. The robot had to complete 10kg box handling, seven-part picking, and sub-millimeter assembly of 18 engine valves within 20 minutes.
The article's core argument: industrial embodied AI competition is shifting from "whose model is stronger" to "who can make models actually deliver." Lingxi's approach pairs the CONWAY industrial-native model with the ROSS Harness execution engine: the model handles local action decisions (force-tactile enhancement, single-step policies, latent action alignment, lightweight safety layers—each backed by papers), while the Harness handles task decomposition, skill orchestration, anomaly recovery, and layered safety monitoring, accumulating real executions into skills, memory, and training data.
The team proposes "effective output = success rate × throughput," claiming ROSS Harness delivers a 3.6× improvement. Rather than chasing the hardest demos, the company targets flexible, high-mix low-volume stations like CNC machine tending, betting that "models can be swapped while the Harness endures" is the key to scaling. CTO Duan Yifan's doctoral advisor, Prof. Ji Jianmin, joined as co-founder and chief scientist.
More from Embodied
- AGIBOT robot lands flying side kicks and sword fights on camera; claims 44% of humanoid market — heyshrutimishra · 2026-08-28
- Skild vs GEN: Both Call It In-Context Learning, Very Different Paths — gan_chuang · 2026-08-28
- XPeng Projects Iron Humanoid Robot Margin Above 50%, Surpassing Cars — shashib · 2026-08-28
- Apple launches M6 Mac mini: local LLMs run up to 4.8x faster than M4 — FinanceYF5 · 2026-08-28
- Robotics startup Microduck surpasses $1M in sales — mishig25 · 2026-08-28
- First interview of humanoid robot running an LLM claims founder — Scobleizer · 2026-08-28