Microsoft Research Pushes AI Into the Physical World With Robots Trained Using Simulation and Real Data

Microsoft is not stopping at digital assistants but is now moving AI into the real world by developing what it calls AI for the physical world which mainly involves robots that can learn, change their behaviour and still operate under the varying conditions. Recently, Microsoft has been releasing a series of studies in which it wants to show how it has been working towards making robotic systems training especially in the areas where it is challenging to collect training data more effective.

Robotic systems need more than just understanding language. Their requirement is the possession of physical intelligence. The physical intelligence comprises the ability to see what is going on around, to think and then to move safely. This kind of learning demands real-world experience. Generating that experience though is a costly and lengthy process.

The most common technique in building training datasets for robots is by using teleoperation. To put it simply, a robot is controlled by a human from a distance allowing it to do the things and the data thus generated are used for training. Microsoft admits that although this has become a common practice, it is however not always the case that it can be used. There are some places that are too complicated, dangerous or simply not easy to get to.

In order to overcome this shortcoming, the study highlighted in Microsoft’s tale is all about enhancing robot datasets with synthetic demonstrations, which are generated by simulation and reinforcement learning. So, in other words, robots can be taught through virtual tests that incorporate real-life conditions thus providing training data that is more diverse and larger without the need for continuous human oversight.

One of the main elements of this research is teamwork. Microsoft cites a statement from Professor Abhishek Gupta of the University of Washington, who clarifies that the group is alongside Microsoft Research enriching pre-training datasets obtained from real-world robots using these synthetic methods.

The objective is clear yet challenging: create robotic systems that are smarter in terms of generalization. Instead of being good only in one specific environment, robots should be able to deal with different kinds of situations, e.g. different objects, new room layouts, unexpected obstacles, or even changes in lighting and movement.

This trend could have a considerable impact on a wide scope of industries where automation is adopted in the first place: warehousing, manufacturing, healthcare assistance, and home robotics. Furthermore, it is indicative of the changing focus of AI research, where success is no longer solely judged by what a model can express but also by what it can achieve without any safety risks.

Microsoft’s strategy demonstrates that AI’s future does not revolve around interactions confined to the screen. Rather, it is gradually leading to the development of systems that can physically learn from their surroundings and enhance their capabilities in the real world.