Boston Dynamics released the latest demonstration video of its electric Atlas robot, which can crawl, run, and do Thomas jumps, demonstrating its amazing flexibility and anthropomorphic movements, which has attracted widespread attention.Behind this demonstration, it is inseparable from the help of the AI institution RAI Institute. The two parties collaborated to develop an advanced training strategy based on motion capture and reinforcement learning.
Unlike in the past, Atlas' movements this time are not pre-programmed, but implemented through motion capture and AI learning.Real actors put on motion capture suits and complete various actions. These action data are then input into the Atlas system, which is learned and imitated by the AI algorithm.
In order to accelerate the learning process of Atlas, RAI Institute uses a simulator driven by a physics engine. The simulator can generate massive training data and simulate various action scenarios, allowing Atlas to be trained efficiently in a virtual environment.
The training data for each action comes from approximately 150 million simulator runs.These "skills" mastered in the virtual world can be directly transferred to the physical Atlas robot without additional training, greatly improving training efficiency.
Last year, Boston Dynamics released the video "Goodbye, Hydraulic Atlas", officially announcing the retirement of Atlas, the hydraulically driven bipedal humanoid robot it developed. The switch from hydraulic drive to all-electric is an important step towards practical application of the Atlas robot.
Electric systems have many advantages over hydraulic systems: they are quieter, more responsive, and easier to maintain.This also enables finer motion control, which is crucial for tasks requiring high precision and agility. In addition, electric systems are more energy-efficient and environmentally friendly, helping to reduce operating costs and environmental impact.