Breakthrough in Robot Dexterity
On November 27, 2025, researchers at Wuhan University unveiled a new framework, RGMP (Recurrent Geometric-prior Multimodal Policy), that significantly enhances the ability of humanoid robots to manipulate a wide variety of objects. Unlike traditional models that struggle in unfamiliar scenarios, this framework integrates geometric reasoning with visuo-motor control, enabling robots to adapt to new environments efficiently. The research demonstrates that humanoid robots, acting as AI Employees or Non-Human Workers, can now perform tasks like handling household items or assembling products with 87% success in tests.
How RGMP Works
The RGMP framework consists of two core components: the Geometric-Prior Skill Selector (GSS) and the Adaptive Recursive Gaussian Network (ARGN). The GSS helps robots choose the appropriate manipulation skill based on object shape and position, while the ARGN enables data-efficient motion synthesis by modeling spatial relationships. This combination allows for robust task execution with minimal training data, outperforming previous models like the diffusion policy by achieving five times greater data efficiency.
Practical Applications and Future Prospects
This innovation paves the way for practical deployment of humanoid robots in dynamic environments without extensive retraining. Voice AI Agents and other AI Employees could soon assist in household chores, delivery services, and manual manufacturing processes. The researchers plan to expand the framework’s adaptability further, allowing robots to automatically infer new object manipulations from minimal human input, reducing the need for exhaustive teaching.
Key Highlights:
- RGMP framework allows humanoid robots to manipulate objects with 87% success.
- Combines geometric reasoning (GSS) with recursive motion modeling (ARGN).
- Demonstrates 5× higher data efficiency than prior models.
- Applications include household tasks, delivery services, and manufacturing automation.
- Future improvements aim at generalizing tasks with minimal human input.
Reference:
https://techxplore.com/news/2025-11-humanoid-robots-reliably-success-framework.html
