musk\'s openai training robots to think like humans in virtual reality (videos)

by:Marslite     2019-10-11
Musk\'s latest joint venture, OpenAI, is a non-
A profitable organization with $1 billion in funding aims to create \"friendly\" artificial intelligence (AGI).
Ai has pre-planned the results and made decisions based on specific inputs.
This is why ordinary artificial intelligence is considered \"narrow\" or \"weak\" within the industry, because the goals it can achieve are very limited.
On the other hand, AI uses input more widely in the same way as humans, making mistakes until the right results are realized and emphasized through reinforcement.
This constitutes \"real\" or \"powerful\" AI.
OpenAI\'s team has trained a self.
Learning algorithms in virtual reality (VR)
Carry out a task, although nothing crazy yet;
Just stack colored blocks in a series of towers.
We developed and deployed a new algorithm.
In a blog post on Tuesday, OpenAI wrote: \"Shooting imitation learning allows humans to communicate how to accomplish new tasks by performing new tasks in virtual reality . \".
The algorithm is driven by two kinds of neural networks: The Visual Network and the imitation network, both of which imitate the process in the human brain.
The visual network contains thousands of analog images that combine a wide variety of shapes, sizes, colors and textures in addition to the ambient lighting effects.
The visual system has never trained with real images because it\'s too time consuming and laborious --
It is intensive for the research team.
This also means that the visual system does not rely on copying the exact scene it has seen before, but can react to a wide variety of situations.
The team added the algorithm \"observe the demo, process the demo to infer the intent of the task, and then complete the intent from another configuration\" in its blog \".
The way algorithms are learned is called \"one-
In this kind of simulation learning, humans demonstrate tasks in virtual reality, and learning algorithms immediately copy tasks in the physical world from any starting point.
While this is not yet a technical miracle, it all has to start somewhere, and these \"baby steps\" are only the first in the long and exhaustive process of educating AGIs to help humanity in the real world.
\"Babies are born with the ability to imitate the behavior of others,\" says Josh Tobin, a technician at OpenAI . \".
\"Imitation allows humans to learn new behaviors quickly.
We also hope that our robots can learn in this way.
\"With a mission demo, we can replicate it under many different initial conditions.
It only takes an extra demo to teach robots how to build different block permutations.
Read more: Elon Musk\'s neural network can represent the next stage of human evolution. The research team also realized that in order to make the algorithm effective in reality
It needs to manage incomplete data and imperfect scenarios.
To achieve this, they introduce \"noise\" in the code or \"policy\" of the control algorithm, so that when things go wrong, it is forced to adapt, and they always do so.
\"Our robot has now learned to perform tasks, although its actions must be different from those in the demo,\" Tobin explained . \".
\"With a mission demo, we can replicate it under many different initial conditions.
It only takes an extra demo to teach robots how to build different block permutations.
\"The value of this adaptability cannot be underestimated because training in virtual reality allows researchers to create any number of potential environments for future robots to develop multiple skills.
Examples can be carried out from search and rescue operations after the earthquake, in nuclear leak areas such as Fukushima, and even on the rocky terrain on Mars, to repair the damaged rover.
Custom message
Chat Online 编辑模式下无法使用
Chat Online inputting...