Zhiyuan's Yao Maoqing Says General Large Models Struggle with Physical Operations; Embodied Foundation Models Will Move Toward Layered Architecture

Recently, discussions around general-purpose large models and specialized embodied intelligence models have continued to heat up. As general foundation models led by giants such as OpenAI keep penetrating into the physical world and the capability boundaries of general large models further expand, is there still a need to develop embodied models trained specifically for robots?

"General large models have significant shortcomings in real physical robot operation, and future systems will inevitably move toward a layered architecture," Yao Maoqing, co-founder, co-president, and president of the embodied business system at AgiBot, gave a clear judgment in an interview with a Securities Times reporter.

Yao Maoqing pointed out that general large models demonstrate closed-loop task capabilities in simulation benchmarks, but in the physical world their shortcomings are very obvious. He listed two core pain points of general large models in the embodied intelligence field: first, poor execution efficiency and huge Token consumption. A physical operation that a human can complete in tens of seconds to a minute may take a general large model 1–2 hours while consuming massive amounts of Tokens, which simply cannot meet the takt requirements of real production lines.

Second is the lack of low-level physical control capability. General large models are good at high-level reasoning such as generating CAD and 3D models, but when facing practical tasks involving highly dynamic motion and force-tactile feedback control, their measured success rate is almost zero, because they have never learned low-level physical interaction.

Based on this, Yao Maoqing believes that embodied intelligence and general large models are not in a replacement relationship, but will move toward a "layered architecture." In this architecture, general large models will increasingly assume the role of System 2 (high-level long-horizon planning), acting as an Agent that calls various skill tools; the task skills output by world models and VLA (vision-language-action models) are the middle-layer tools being called; and at the very bottom is the 100–1000Hz high-frequency motor real-time control layer, which is deeply coupled with the body and tasks and must be continuously iterated by embodied intelligence companies.

Yao Maoqing said that even if general large models want to achieve expert-level performance in the physical world, they likewise need massive high-quality physical post-training data, and every layer requires continuous investment and construction by embodied companies.

At the 2026 AgiBot Innovation Day, Yao Maoqing released the new-generation embodied foundation model GO-3 Preview on site and announced that the model will be fully open-sourced by the end of 2026. Yao Maoqing told reporters that GO-3 Preview adopts a brand-new WRAM (World Reasoning Action Model) native architecture. In real deployment scenarios, the body, data, and model are highly coupled, and the entire chain is subject to the "bucket effect," meaning that if any link is missing, the business loop cannot be closed.

Data released at the Embodied Operation Productivity Sub-forum of the 2026 AgiBot Innovation Day show that in training data scaling experiments, the Zero-Shot average success rate of GO-3 Preview on 100 tasks across 20 skill categories increased from 13.4% to 56.4%; in model scaling experiments, the zero-shot success rate increased from 56.2% to 73.2%.

"If a robot manufacturer can only do single points, it cannot form a business loop," Yao Maoqing said. AgiBot will use a full-stack technology system to form a complete physical AI technology loop across three stages: pre-training, post-training, and continuous learning after deployment, so that in real scenarios robots can truly turn technology into measurable productivity value.