Recently, Tencent’s self-developed multi-mode four-legged robot “Robot dog Max” has been upgraded again. According to Tencent officials, the upgrade is to “apply pre-training and reinforcement learning technology to the field of robot control”, that is, to conduct representation learning on common animal data sets, and store the learned potential expressions in a deep neural network, so that its actions and behaviors are more similar to real animals.

▲ Figure source Tencent, the same below

According to reports, Tencent robot dog Max learns the real dog’s walking, running, jumping, standing and other actions, and will flexibly use these gestures to solve various tasks with obstacles, such as creeping forward, hurdling, and running between obstacles.

DSDX452 5716075-P

DSDX452 5716075-P

IT Home learned from Tencent that Tencent specially invited a dog to assist the robot dog in the collection of action data, and collected a certain number of regular motion posture data of real dogs on the flat through the dynamic capture technology of Tencent games, and then used redirection technology to make the data accurately and efficiently mapped to the robot body through the simulation engine.

Subsequently, the robot dog uses these data to build imitation learning tasks in its simulator, and then senses its own joint state in the quasi-real world, exercises on the flat ground according to random instructions, ADAPTS to various movements, and adds complex environmental factors to deal with various obstacles caused by the external environment.

Tencent said that the entire process is trained in the virtual world, without the need for real machine training and adjustment. The robot dog Max will continue to evolve and adapt in the future, and is expected to be put into emergency work such as search and rescue.