In the future, artificial intelligence could be used to train everything from robots to animals and even some basic human tasks. For example, Google’s DeepMind trained various reinforcement learning agents in parallel for 400,000 steps and then evaluated the most promising on a real robot. One of the tasks involved the precise insertion of a USB key stick a computer port.
The agent was provided with reward sketches from over 100 demonstrators and it managed to reach over 80% success rate within 8 hours. Observations came from three cameras located around a cage, as well as two wide-angle cameras and one depth camera mounted at the wrist and proprioceptive sensors in the arm. This setup collected 400 hours of multiple-camera videos of proprioception (perception or awareness of position and movement).
- Slim Design
- Software for device management and backup with password protection. (Download and installation required. Terms and conditions apply. User account...
- 256-bit AES hardware encryption
[Our] approach makes it possible to scale up RL in robotics, as we no longer need to run the robot for each step of learning. We show that the trained batch [reinforcement learning] agents, when deployed in real robots, can perform a variety of challenging tasks involving multiple interactions among rigid or deformable objects. Moreover, they display a significant degree of robustness and generalization. In some cases, they even outperform human teleoperators,” said the coauthors of the white paper.