8″ height=”1365″ srcset=”https://s3.ifanr.com/wp-content/uploads/2019/10/merlin_162679119_3d7f1386-00bf-45db-b682-bbb2fd35b93a-superJumbo.jpg!720 2048w, https:/ /s3.ifanr.com/wp-content/uploads/2019/10/merlin_162679119_3d7f1386-00bf-45db-b682-bbb2fd35b93a-superJumbo-360×240.jpg 360w, https://s3.ifanr.com/wp-content/uploads/ 2019/10/merlin_162679119_3d7f1386-00bf-45db-b682-bbb2fd35b93a-superJumbo-768×512.jpg 768w, https://s3.ifanr.com/wp-content/uploads/2019/10/merlin_162679119_3d7f1386-00bf-45db-b682-bbb2fd35b93a -superJumbo-1024×683.jpg 1024w” sizes=”(max-width: 2048px) 100vw, 2048px”>
Peter Welinder, Head of Robotics at OpenAI represents, in fact, many robots can easily solve the Rubik’s Cube, but the biggest difference between Dactyl and these robots is that those robots are made to solve the Rubik’s Cube, but Dactyl can do more. task.
We are trying to build a generic robot that can perform multiple operations like a human hand, rather than being limited to a specific task.
So, what Dactyl really does is not the one-handed solution, but the process of learning this skill. Because in the whole process, the researchers did not specifically program the operation of the robot, everything depends onDactyl comprehends himself.
In order to make robots “self-taught”, artificial intelligence is indispensable. Dactyl uses a deep learning model in a virtual environment. The training mode of this virtual environment has the advantage that it does not consume real-world time, and there is no need to worry about the robot breaking or hurting others during training. .
Dactyl has accumulated tens of thousands of years of training experience in the virtual world, but in reality only a few months, quite a bit of “the mountain is one day, the world has been a thousand years” feeling, this training method has been greatly shortened The learning time of AI requires thousands of ultra-high-performance CPUs and GPUs to run simultaneously.
Dactyl through this training can also cope with various emergencies. For example, in the process of solving the Rubik’s Cube, the researchers continue to poke it with some objects, and also use paper dust and foam to interfere, but Dactyl can still The task was completed, and this situation was not simulated in the training.
