According to Tencent officials, this upgrade applies “pre-training and reinforcement learning technology to the field of robot control”, that is, conducting representation learning on common animal datasets, storing learned potential expressions in deep neural networks, making its movements and behaviors more similar to real animals.
According to the introduction, Tencent‘s robot dog Max has learned real dog movements such as walking, running, jumping and standing. It will flexibly use these postures to solve various tasks with obstacles, such as crawling forward, hurdling, parkouring between obstacles and so on.
Tencent specially invited a dog to assist the robotic dog in collecting motion data. Using Tencent‘s motion capture technology, a certain amount of real dogs’ regular movement posture data on flat ground was collected. Then, using redirection technology, the data was accurately and efficiently mapped onto the robot body through simulation engine.
Subsequently, the robotic dog utilizes this data to construct imitation learning tasks in its simulator and perceives its own joint status in the simulated world. It practices and adapts various movements on flat ground according to random instructions, and incorporates complex environmental factors to cope with various obstacles caused by external environments.
Tencent stated that the entire process is conducted through virtual training without requiring actual machine training or adjustment. The robotic dog Max will continue to evolve and adapt in the future, with hopes of being deployed for emergency tasks such as search and rescue.
Sign up today for 5 free articles monthly!