The result: humanoid robots are now all over Berkeley.
From @TheRundownAI
A team of researchers at the University of California, Berkeley, has created a two-legged robot Cassie that can learn to walk on its own through AI-reinforcement learning, rather than directly programming or imitating.
1. AI simulation training to learn to walk, no need to fall and start over like a baby
The bipedal robot Cassie resembles the shape of our legs, so it is easier to enter urban environments designed for humans than other robots.
To help Cassie learn to walk independently like a human, the research team imagined the process as a baby learning to walk. Although babies do not learn to walk standing directly, they can remember the steps of walking through standing, falling, and stride, and finally learn to walk upright.
2. Strengthen learning in the gait library to make the pace more flexible and steady
The research team based on reinforcement learning (Reinforcement Learning; RL) method, hoping to let Cassie learn to walk more agilely through the systematic learning method. Reinforcement learning, also known as reinforcement learning and evaluative learning, is the way agents learn through “trial and error” to achieve specific goals in environmental interaction.
Prior to this, researchers often controlled bipedal robots to walk through mechanical modeling, but this method was difficult to model complex ground, and robots lacked the ability to adapt to environmental changes and movement stability.
3. AI tracks the walking environment, and the walking speed and altitude are automatically adjusted
Based on the RL method, the researchers established an adaptive speed control walking controller.
This controller can track Cassie’s walking environment through AI and give her an appropriate walking strategy.
Conclusion: AI-reinforcement learning helps robots move more agilely
Based on the reference movement of the gait library, AI reinforcement learning can help bipedal robots learn walking, turning, squatting and other movement states, and track their walking environment to achieve automatic speed adjustment, turning and other functions, so that the robot can better achieve flexibility and robustness in movement.
In the future, AI reinforcement learning will also help bipedal robots and other robots learn more dynamic and agile behaviors on this basis, helping them cope freely in complex unknown environments.