New AWS Service Lets Any User Create Simulated Worlds To Train Robots, ML and AI

Practitioners in the closely related domains of robotics and machine learning have long had a problem when it comes to fielding real-world solutions. Systems need to be trained to operate in environments that are not known yet. It is impossible to know every future environment, so you can see how hard this is. The state of the art for most is to base solutions on what your best judgement says will be right, use as many lessons learned from previous work you can, and watch the solution closely in execution for the failure scenarios you know are coming.

Let me give you a very simple example. Consider a factory robot that discovers a new wall has been placed on its normal path. Today that probably means a human will need to program a new path. Or consider an imagery system that is designed to detect and count objects and classify them by type. What does it do when it gets an object that sort of looks like the object? Or what if it sees an object that is intentionally camouflaged? How will it perform? The current state of the art is to design systems that overwhelm humans with the need to inform the solution with their judgement.

And in both the examples of robots and machine learning like image classification systems, it can be very hard to check expected accuracy in test situations before they are fielded. So there are two big categories of problems here, getting lots of good training data, and then testing systems before fielding.

Which leads us to the significance of the news announced by AWS today. They just launched a very cool robotics and autonomous system simulation and testing tool, called RoboMaker World Forge.

The way this works is that a user (any user, you don’t have to be a developer to do this) can click a few buttons and make some selections and then create multiple simulated environments. For example, create 50 different indoor office 3D environments with different furniture placement and different flooring, different layouts etc. Then these 50 different environments can be used as training data for the robotics or ML solution. Or they can be used for testing.

Right now the current version is focused on indoor environments. But I have seen a version that is being used by NASA to test ML on for the Mars Rover.

The current capability is incredible, but coming soon, really soon, is the ability to dynamically create training and simulation data for outdoor spaces (cities, open land, deserts, ocean floor, other planets). The magical thing here is not that this can all be created, but that it can be created in ways that enable AI training and simulation and testing. This is awesome. And will be very disruptive to the market.

Here is the announcement video:

For more see:

, , ,