Scientists have presented a robot that can balance and maintain balance even in difficult conditions that change during the experiment. To do this, they trained AI on a treadmill and skateboard.
The AI developers noted that they created a framework for controlling robots on four legs. It adapts better than more traditional robotic motion control models. To showcase new functionality that adjusts to environmental conditions in real-time, the researchers showed how the device glides on surfaces, skates, and runs on a treadmill with an incline.
“Our design teaches a controller that can adapt to changes in the environment while driving. These may be new scenarios that we did not study during the training. This makes the controller 85% more energy-efficient and more reliable than traditional methods, the researchers note. “During inference, the high-level controller only needs to evaluate a small multi-layer neural network, it does not need control and predictive model (MPC) that would be required to optimize long-term performance”.
The model learns how to move using a treadmill, which consists of two belts – their speed changes independently of each other, but the robot still maintains balance. This simulation training is then carried over to the Laikago robot in the real world. Researchers have released a special video about simulations and lab work to popularize the technology.
AI experts from Nvidia, the University of California, the University of Texas at Austin, and the University of Toronto participated in this study. Their design includes a high-level controller using learning amplification and a lower-level controller based on an AI model.