Scientists presented a robot that knows how to balance and maintain equilibrium even under difficult conditions that change during the experiment. For this purpose, they trained the AI on a treadmill and skateboard.
The AI developers noted that they created the basis for controlling the robots on four legs. It adapts better than more traditional robot movement control models. To show the new functionality, which adapts to the environment in real time, the researchers showed how the device slides on the surfaces, skateboarding and running on a treadmill.
“Our development teaches a controller that can adapt to changes in the environment as it moves. These may be new scenarios that we didn’t learn during the training. This makes the controller 85% more energy efficient and reliable than traditional methods,” the researchers note. – During the output, a high tier controller should only evaluate a small multi-layer neural network, it doesn’t need the control and prediction model (MPC) that might be needed to optimize long-term performance.
The model is trained to move using a treadmill, which consists of two ribbons – their speed varies independently of each other, but the robot still maintains its balance. This simulation training is then transferred to the Laikago robot in the real world. The researchers released a special video about simulations and laboratory work to promote the technology.
This study involved AI experts from Nvidia, the University of California, the University of Texas at Austin and the University of Toronto. Their development includes a high level controller that uses learning amplification and a lower level controller based on the AI model.