For a roboticist, reproducing the movements of one hand is a puzzle. To better solve it, North American researchers have created a robotic hand that can improve its dexterity without human intervention, thanks to a machine learning algorithm.
In robotics , the hand is one of the most complex parts to master. It is necessary to combine a high degree of mobility, a reactivity and a precision likely to approach the performances of a human hand. Great progress has been made in terms of dexterity, speed (including a robot hand unbeatable with chifoumi) or even sensitivity to touch.
Many simple gestures, which we perform daily with our hands, remain real technical challenges for roboticists. However, nothing is impossible, as just demonstrated by a team from the University of Washington (University of Washington, UW). Its researchers designed a robotic hand with five fingers that shows great skill. Above all, thanks to an artificial intelligence (AI), it is able to improve its dexterity without human intervention.
To design this Adroit Manipulation Platform, UW experts had to face several challenges. Starting with the design of a robotic hand capable of offering flexibility and responsiveness capable of imitating a human hand. For this, the researchers left on the basis of the Shadow Dexterous Hand, an artificial hand model manufactured by the British company The Shadow Robot Company.
She operates all her joints in less than three seconds
Animated by a pneumatic system which actuates 40 tendons, it has a size and a shape comparable to those of a human hand and reproduces all the degrees of freedom. The version designed by the UW team is extremely fast. In this demo video available on YouTube, it only takes 2.44 seconds to operate all of its joints one after the other.
Then, the researchers developed a machine learning algorithm allowing to model complex hand movements, involving the five fingers, and to determine the most suitable combination according to the tasks to be accomplished such as typing on a keyboard. or drop a stick and catch it in the air.
After being tested on a computer, this AI was associated with the artificial hand to make it work in real conditions. The operation is as follows: the system collects the data emitted by 130 sensors and motion capture cameras then inserts them into the machine learning algorithm which will thus refine its existing models. The video shows how, without outside intervention, the hand becomes more and more able to twirl between its fingers a plastic tube filled with coffee beans.
The next step will be to develop a more extensive learning capacity so that the robotic hand can manage objects and handling scenarios of which it has no prior knowledge.
This innovation can potentially advance robotics in many fields such as industry, handling, surgery or personal services. The main obstacle is the high manufacturing cost of this hand, which is around 300,000 dollars (more than 263,000 euros at current prices). For the moment, the UW team intends to focus on the development of its platform and is not talking about a commercial project.
