At the WAIC 2026 Embodied Intelligence Forum, Yao highlighted scalable data infrastructure and end-to-end learning systems as key foundations for developing general-purpose Physical AI.
-- On July 17, 2026, Maniformer today announced that its Chairman and CEO, Maoqing Yao, presented the company’s approach to addressing three key challenges in Physical AI—data scarcity, physical-world representation, and closed-loop learning—at the WAIC 2026 Embodied Intelligence Forum. During his presentation, Yao highlighted scalable data infrastructure and end-to-end learning systems as important enablers for the development and real-world deployment of general-purpose Physical AI.
Co-hosted by AGIBOT and Maniformer, the forum brought together industry participants to discuss advances in embodied intelligence. Yao, who also serves as Partner, Senior Vice President, and President of the Embodied Intelligence Business Unit at AGIBOT, focused his remarks on the technical and data challenges affecting the development of Physical AI.
Three Barriers Slowing Physical AI
According to Yao, the next breakthrough in Physical AI depends on overcoming three fundamental challenges:
- Data bottlenecks, caused by the scarcity of large-scale, high-quality real-world interaction data.
- Representation challenges, which make it difficult for AI systems to accurately model and understand the physical world.
- Closed-loop learning limitations, preventing robots from continuously improving through real-world deployment.
"Scaling models alone is not enough," Yao noted. "Physical AI requires a complete learning loop that connects data, models, simulation, deployment, and continuous optimization."
Building an End-to-End Learning Framework
To address these challenges, AGIBOT has developed an end-to-end technology framework spanning pre-training, mid-training, and scalable online post-training to optimize its Vision-Language-Action (VLA) models for generalist robots. This closed-loop system continually trains models across robot fleets to adapt to real-world environments while retaining general capabilities.
Rather than relying on a single model architecture, the company's roadmap advances two complementary directions in parallel: Vision-Language-Action (VLA) models for embodied decision-making and World Action Models (WAM) for modeling physical interactions. These efforts are designed to converge into a unified World Reasoning Action Model (WRAM) that integrates perception, reasoning, and action within a single foundation model.
Supporting this architecture is a large-scale training infrastructure that combines simulation, reinforcement learning, and distributed real-robot training, enabling continuous iteration across both virtual and physical environments.
Data as the Foundation of Physical AI
Yao emphasized that advances in model capability must be matched by advances in data infrastructure.
"Without scalable, high-quality physical-world data, even the most advanced models will struggle to generalize in real-world environments," he said.
To support large-scale Physical AI development, Maniformer has built an end-to-end data infrastructure spanning data acquisition, governance, management, and deployment.
The ecosystem includes the MEgo series of data collection devices, the MEgo Engine data governance platform, and the AGIBOT WORLD real-world robot dataset. Together, these components form a layered data supply pyramid and a deployment feedback flywheel, allowing operational robot data to continuously improve future model training.
This integrated model-and-data flywheel has already been validated in production environments, including large-scale deployment on consumer electronics manufacturing lines, where robotic task execution achieved up to a 99.99% task success rate in selected production scenarios.
"Models determine where intelligence begins, but data determines where it can ultimately go," Yao said. "Only through a continuously evolving model-and-data flywheel can Physical AI overcome today's barriers and move toward general-purpose intelligence."
As Physical AI continues to evolve from research to large-scale commercial deployment, Yao believes the industry's next breakthroughs will depend not only on stronger foundation models, but also on scalable data infrastructure capable of continuously improving intelligence through real-world experience.
About Maniformer
Maniformer is a world-leading Physical AI data infrastructure company providing end-to-end solutions for real-world data acquisition, governance, management, and deployment. Through its scalable data platform and global ecosystem, Maniformer helps organizations develop, deploy, and continuously improve intelligent robotic systems.
Contact Info:
Name: Linko Song
Email: Send Email
Organization: Maniformer
Website: https://maniformer.ai
Release ID: 89198331
In the event of any inaccuracies, problems, or queries arising from the content shared in this press release, we encourage you to notify us immediately at error@releasecontact.com (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). Our diligent team will be readily available to respond and take swift action within 8 hours to rectify any identified issues or assist with removal requests. Ensuring the provision of high-quality and precise information is paramount to us.
