If you collect data only on real robots, you are limited to hours per

[rifat28dddd]

We can use this data to learn how people interact with objects and how objects behave in different situations, which provides a common sense prior for the basic robot model. However, internet data does not have action signals, and we cannot retrieve the robot’s motion control signals from the internet. This leads to another data strategy, which is to use simulations. In a simulation, you have all your actions and you can observe the results of those actions in a specific environment.



The advantage of simulation is that it is essentially unlimited cameroon phone numbers data, and the amount of data increases as computing resources increase. The more GPUs you invest, the more data you generate, and the data is in real time. day. With GPU-accelerated simulators, we can speed up real time by more than . times. We are able to collect data at a much higher rate during the same working hours.



But the weakness of simulation is that no matter how good the graphics pipeline is, there will always be a gap between simulation and reality. The laws of physics and visual effects are different from the real world,and the content is not as diverse as the situations we encounter in the real world. Finally, there is real robot data, which does not have the gap between simulation and reality because it is collected on real robots, but it is much more expensive to collect because people have to be hired to operate the robots.