Zhiyuan's Yao Maoqing: Embodied AI Will Adopt a Layered Architecture as General Models Struggle with Physical Tasks
通用大模型难解物理实操？智元姚卯青：具身基座将走向分层架构
Yao Maoqing, co-founder and president of Zhiyuan, stated that general-purpose large models lack the efficiency and low-level physical control required for real-world robotics. He proposed a layered architecture where general models handle high-level planning, while specialized embodied systems manage mid-level skills and low-frequency motor control. At the 2026 Zhiyuan Innovation Day, Yao unveiled the GO-3 Preview model, which uses a new WRAM architecture. The company plans to fully open-source the model by the end of 2026.
Source: Chinese robotics — Robot learning research · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
13.4%
Reported zero-shot result; the excerpt does not specify an evaluation protocol.
View original evidence
2026智元创新日具身作业生产力分论坛公布的数据显示，在训练数据规模扩展实验中，GO-3 Preview在20类技能100个任务上的Zero-Shot（零样本）平均成功率由13.4%提升至56.4%；在模型规模扩展实验中，零样本成功率由56.2%提升至73.2%。
Open source S16Reported success rate
56.4%
Reported zero-shot result; the excerpt does not specify an evaluation protocol.
View original evidence
2026智元创新日具身作业生产力分论坛公布的数据显示，在训练数据规模扩展实验中，GO-3 Preview在20类技能100个任务上的Zero-Shot（零样本）平均成功率由13.4%提升至56.4%；在模型规模扩展实验中，零样本成功率由56.2%提升至73.2%。
Open source S16Reported success rate
56.2%
Reported zero-shot result; the excerpt does not specify an evaluation protocol.
View original evidence
2026智元创新日具身作业生产力分论坛公布的数据显示，在训练数据规模扩展实验中，GO-3 Preview在20类技能100个任务上的Zero-Shot（零样本）平均成功率由13.4%提升至56.4%；在模型规模扩展实验中，零样本成功率由56.2%提升至73.2%。
Open source S16Reported success rate
73.2%
Reported zero-shot result; the excerpt does not specify an evaluation protocol.
View original evidence
2026智元创新日具身作业生产力分论坛公布的数据显示，在训练数据规模扩展实验中，GO-3 Preview在20类技能100个任务上的Zero-Shot（零样本）平均成功率由13.4%提升至56.4%；在模型规模扩展实验中，零样本成功率由56.2%提升至73.2%。
Open source S16task_duration
1–2 hours
View original evidence
人类几十秒到一分钟就能完成的物理操作，通用大模型可能需要耗费1—2小时，同时消耗海量Token，这根本无法满足真实产线的节拍要求。
Open source S8
- code: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
2026智元创新日具身作业生产力分论坛公布的数据显示，在训练数据规模扩展实验中，GO-3 Preview在20类技能100个任务上的Zero-Shot（零样本）平均成功率由13.4%提升至56.4%；在模型规模扩展实验中，零样本成功率由56.2%提升至73.2%。
Open source S16
人类几十秒到一分钟就能完成的物理操作，通用大模型可能需要耗费1—2小时，同时消耗海量Token，这根本无法满足真实产线的节拍要求。
Open source S8
在2026智元创新日上，姚卯青现场发布新一代具身基座模型GO-3 Preview，并宣布该模型将于2026年底全面开源。
Open source S14
Source：Chinese robotics — Robot learning research · finance.sina.com.cn