Recently, Chinese research teams have made significant breakthroughs in the field of robot algorithms, proposing the world's first unified theory of "force position hybrid control algorithm". This algorithm does not rely on force sensors, allowing the robot to learn both position and force control simultaneously. The success rate of related tasks has increased by about 39.5% compared to the strategy that only uses position control. More noteworthy is that its related papers have currently won the Outstanding Paper Award at the International Robot Learning Conference, which is the first time since the establishment of the award that a team of Chinese scholars has won it.

At the experimental site of Beijing General Artificial Intelligence Research Institute, a quadruped mechanical dog equipped with this new algorithm is carrying out the training task of wiping whiteboards in an orderly manner. Researchers used this practical scenario to explain to reporters the core principles and prominent advantages of the "force position hybrid control algorithm".

Jia Baoxiong, a researcher at Beijing General Artificial Intelligence Research Institute: The mainstream visual language action models actually take photos taken by cameras as input and then make trajectory predictions. This trajectory may cause the robot to not be able to strictly adhere to the whiteboard during the process of wiping, resulting in the inability to wipe this situation. In the force position mixed control model, we try to fill in the dimension of force as much as possible, achieving the effect of applying pressure on the whiteboard on the basis of precise positioning, so that it can closely fit the whiteboard and erase the words on it. That is to say, we not only flattened the dimension of force on the original visual language action model, but also achieved the effect of simultaneously mixing force and position control.

According to researchers, the widely used visual language action model often falls short in dealing with many tasks in real life. The core issue lies in the fact that most of these tasks involve complex contact scenarios, such as when wiping a blackboard, the robotic arm must adhere to the surface while maintaining appropriate pressure, the switch cabinet door needs to accurately perceive the internal push-pull spring structure, and the robot not only needs to "go where it goes and extend its hand where it goes", but also needs to "understand how much force should be used". Before the 'force position hybrid control algorithm', all of these needed to be solved through force sensors.

Jia Baoxiong, a researcher at Beijing General Artificial Intelligence Research Institute: Force sensors are actually more commonly added to fixed robotic arms, making installation more complicated and expensive. The force position mixed control model has improved the success rate of these tasks requiring force by nearly 40% compared to the traditional VLA model (visual language action model).