They develop a system to train quadrupedal robots in the domain of soccer balls
Quadrupedal robots have aroused great interest within robotics research environments, given the wide spectrum of possibilities of use associated with these systems.
Expanding on the possibilities explored so far with these specimens, new research has managed to solve a previously unattainable challenge: the possibility of training quadrupedal robots to control soccer balls.
A new training system for robots, to equip them with the ability to master balls
Machine learning models that are typically built into robots of this class to interact with their environment are difficult to adapt to situations where their dynamics are difficult to model, such as a deformable soccer ball.
A recent article published on arXiv.org presents, with the aim of achieving precise shooting skills, a training system for quadrupedal reinforcement learning robots, based on two key elements: a movement control policy and a planning policy.
As for the motion control policy, its role lies in learning various full-body movements, to track random parametric trajectories of the toes while maintaining balance while standing. Regarding the motion planning policy, your responsibility lies in shooting the deformable soccer ball to the desired target.
“Developing algorithms to enable a legged robot to shoot a soccer ball at a given target is a challenging problem that combines robot motion control and planning in a single task,” says part of the report’s presentation. this project.
To solve the described difficulty, the team behind this research considered the dynamic limitation and stability of movement during the control of a dynamic legged robot. In addition, the planning of the movement to shoot the deformable ball was contemplated, a difficult data to model, since the ball rolls on the ground with an uncertain friction to a desired location.
The experiments carried out at a practical level with a robot, whose development was documented through a videowere able to validate the proposed methodology, by demonstrating the feasibility of achieving a firm control policy and an accurate planning policy that, in this case, allowed a “robot dog” to assume the role of “soccer player”.
