A research group on Thursday released Genesis, an artificial intelligence simulation engine designed to ease robot development. 

The group included more than 50 researchers from about a dozen universities. Nvidia Corp. contributed as well along with the MIT-IBM Watson AI Lab, an industrial-academic AI institute staffed by MIT and IBM Corp. experts.

Robots often rely on AI models to navigate their surroundings or perform actions. Those AI models, in turn, are trained through a trial-and-error process. A neural network is given a goal, such as steering a robot from one part of a warehouse to another, and repeats the task until it learns to perform it effectively. 

Carrying out such AI training in the real world can be prohibitively difficult. The process requires pricey robotics hardware that isn’t always readily available for researchers. Additionally, the large number of attempts an AI model requires to learn a task means that achieving the necessary accuracy can take months. 

Researchers address the challenge by training robots’ AI models in virtual environments. Simulations can run significantly faster than a real-world training session. Moreover, they can be parallelized. This makes it possible to save even more time by running AI training sessions side-by-side rather than one after another.

The new Genesis platform eases the task of creating robot simulations. According to the software’s developers, it can train robotics-focused AI models 10,000 times faster than would be possible in the real world. In practice, that means that a decade worth of training can be compressed into one hour of compute time.

Genesis is built on a physics engine that can simulate a wide range of materials as well as phenomena such as rain. It does so using specialized simulation algorithms, called physics solvers, that each focuses on generating a different set of data points. 

The speed of Genesis simulations is made possible by several low-level optimizations built into the software. 

When virtual objects don’t actively move or change, they require less hardware resources to render than objects that do. Genesis uses a feature called auto-hibernation to reduce the amount of commuting capacity it spends on static objects. Additionally, the researchers have optimized its collision checking mechanism, which is responsible for preventing simulated robots from overlapping when they shouldn’t.

The second component of Genesis is a tool called RoboGen. It enables users to generate robotic arms, robotic vacuums and other kinds of autonomous machines. RoboGen also makes it possible to specify the tasks that those robots should perform, as well as customize related details such as how they should behave if a component malfunctions.

Genesis provides a chatbot interface for designing simulations. Instead of writing code, users can define the configuration of a virtual environment using natural language descriptions. That significantly reduces the amount of time and effort involved in the task. 

Besides teaching robots how to perform new tasks, Genesis also lends itself to producing training data for AI projects. 

The platform allows users to capture footage of their robot simulations. Moreover, it provides the ability to customize the camera angle and apply ray tracing, a rendering method that improves the visual fidelity of 3D models. Researchers can use those features to closely align the videos in their training dataset with the requirements of an AI project. 

Initially, the developers of Genesis are only open-sourcing the platform’s physics engine and RoboGen module for generating virtual robots. The chatbot interface for designing simulations using natural language will follow suit in “the near future.”

Image: Genesis