Researchers at Wuhan University have developed a new framework that allows robots to manipulate objects more easily. Featured in a new paper above arXiv,This approach allows humanoid robots to grasp and handle a wider variety of objects than they currently can. It’s possible now.
Currently, humanoid robots are good at tasks such as using tools, grasping objects, and walking, but they have inherent limitations. In most cases, a task can fail if the shape of the object changes or the lighting changes.
Robots may also struggle to complete tasks they are not specifically trained to do. This lack of generalizability is widely considered to be one of the major limitations of this technology.
To overcome this, the Wuhan team set out to develop what they called the Recurrent Geometric Priority Multimodal Policy, or RGMP for short. This framework is designed to allow humanoid robots to have some kind of built-in common sense about things like shape and space.
It also gives robots a way to better select the skills needed for a task and a more data-efficient way to learn movement patterns.
Further generalization of humanoid robots
The ultimate goal is to help robots choose appropriate actions and adapt to new environments with much less training data than before. According to the team, RGMP consists of two main parts.
The first is called Geometric-Prior Skill Selector (GSS), which helps robots decide which “tools” and skills are best suited for a task. Robots can use GSS, such as with cameras, to understand the shape, size, and orientation of objects.
With this information in hand (so to speak), the robot can determine what it needs to do to complete a given task (i.e. lifting, pushing, grasping, holding with both hands, etc.).
The second one is called an adaptive recurrent Gaussian network (ARGN). Once the robot selects a skill, ARGN actually helps the robot perform the task. This is achieved by modeling the spatial relationships between robots and objects.
It also helps predict behavior step by step and is very data efficient (requiring far fewer training samples than typical deep learning methods).
Combining ARGN and GSS allows robots to better complete tasks without the need for thousands of demonstrations or training. In testing, robots using this framework were able to achieve an impressive 87% success rate on new tasks that the robot had no experience completing.
Dramatic improvement compared to competitors
The team also found that the framework is approximately five times more data efficient than current dissemination policy-based models (currently state-of-the-art). This is impressive and could be very important in the future.
Once robots can reliably manipulate objects without having to be retrained for new situations, they could actually be used for tasks such as cleaning, organizing, and even helping around the house like cooking.
It will also take humanoid robots to the next level when working in locations such as warehouses. restaurants, manufacturing, etc. Looking to the future, the team now wants to extend RGMP to allow robots to learn new tasks with little human guidance.
We also plan to enable RGMP to independently infer the correct behavior for entirely new objects and automatically generate task-specific behavior patterns. “Our future research will focus on enhancing the RGMP framework’s ability to generalize across different tasks,” explains study lead author Xuetao Li.
“We also plan to explore automatic inference of task-specific motion trajectories, which would allow robots to infer new object operations based on minimal human input or prior knowledge, further eliminating the need for intensive instruction in dynamic environments,” he added.
You can see the research for yourself in the journal arXiv.
