Chinese researchers unveil RGMP, a data-efficient AI framework boosting humanoid robots’ grasping skills and generalization.
Researchers in China have introduced a new AI framework designed to enhance humanoid robot manipulation. According to researchers at Wuhan University , RGMP aims to improve grasping accuracy across a broader range of objects and enable robots to perform more complex manual tasks.
Unlike many data-driven methods that rely on large training datasets, RGMP incorporates geometric reasoning to boost generalization in new or unpredictable environments. The framework achieves 87 percent generalization and is 5 times more data-efficient than leading diffusion-based models, combining spatial reasoning with efficient learning. The researchers say the framework could be a step toward more adaptable and capable humanoid systems.Smarter robot skillsFor humanoid robots to operate independently, they must reliably handle multiple objects across different environments. Current machine learning models often work well only when the robot operates in settings similar to those used during training.These systems rely heavily on large datasets and do not fully use geometric reasoning or spatial awareness, making it difficult for robots to adapt in new situations. Vision-language models can understand instructions but often struggle to link them with the correct actions, especially when object shapes or contexts vary. According to researchers, other approaches, like diffusion or imitation learning, require many demonstrations and still fail to generalize. This raises two key questions: how robots can reason about object geometry and how they can learn effectively with fewer examples.The team developed a data-efficient approach that uses geometric reasoning to help robots generalize skills in unseen environments.To address limitations in current robot manipulation systems, the team developed RGMP, a new end-to-end framework that combines geometric reasoning with efficient learning. The first part, called the Geometric-prior Skill Selector , helps the robot choose the correct action based on an object’s shape and task requirements, much as humans decide whether to grasp, pinch, or push. It uses simple geometric rules and works even in new environments. The second part, the Adaptive Recursive Gaussian Network , improves learning from small datasets by storing and updating spatial memory. It models the robot’s interactions with objects over time, thereby avoiding vanishing gradients. Together, these components help robots generalize better and handle more complex tasks with fewer training examples.Efficient robot intelligenceThe team tested the RGMP framework to assess its performance and generalization. Experiments were carried out on two types of robots: a humanoid system and a desktop dual-arm robot equipped with cameras and 6-DoF arms.A dataset of 120 demonstration trajectories was used, and performance was measured through two metrics: selecting the correct skill and executing it accurately. RGMP was compared with leading models, including ResNet50, Diffusion Policy, Octo, OpenVLA, and others. The results show RGMP performed better across multiple manipulation tasks, including unseen objects and new environments. Researchers claim the GSS module improved skill selection by up to 25 percent, while ARGN and Gaussian modeling improved execution accuracy. The system also required far fewer training samples—achieving high performance with just 40 examples, compared to 200 needed by baseline models—demonstrating strong efficiency and adaptability.The team highlights that by linking skills to object context and breaking 6-DoF motions into Gaussian components, the system improves efficiency and generalization. RGMP achieves 87 percent generalization accuracy and uses 5 times less data than the Diffusion Policy during human-robot interaction tests. The results show that integrating symbolic reasoning with learning improves adaptability across new objects and environments. Future research will focus on enabling robots to infer actions for new objects after learning just one example.The Wuhan University team’s research details are available on the arXiv preprint server.
China Humanoid Recurrent Geometric-Prior Multimodal Policy RGMP Robot Robot Training Wuhan University
United States Latest News, United States Headlines
Similar News:You can also read news stories similar to this one that we have collected from other news sources.
Chinese Robot Sets Guinness World Record With 66-Mile WalkThe U.S.-China robot race is getting more literal.
Read more »
All 10 Seasons of This Cult Classic 87% RT Sci-Fi Comedy Series Are Now Free To StreamCrow, Michael Nelson, and Tom Servo in Mystery Science Theater 3000
Read more »
Appliance Warrant Runaround: 87-year-old gets new refrigerator after I-Team gets involvedAn 87-year-old woman says she went without a working refrigerator for almost a month, until she contacted the ABC7 I-Team.
Read more »
Appliance Warranty Runaround: 87-year-old gets new refrigerator after I-Team gets involvedAn 87-year-old woman says she went without a working refrigerator for almost a month, until she contacted the ABC7 I-Team.
Read more »
The 87 Best Black Friday Deals on Running GearWe found the best Black Friday deals on running gear, including men’s and women’s running apparel, shoes, gear, and energy gels.
Read more »
Meet the humanoid robot that just set a new world walking recordThe robot walked over 2.5 full marathons in China.
Read more »