New MIT AI Designs Robots On Its Own

New MIT AI Designs Robots On Its Own

A new MIT project aims to take the guesswork out of robotics. Instead of trial and error to find the right design for a task, you can just ask RoboGrammar. The program just needs to know what parts you’ve got lying around and what you need the robot to do. The team believes RoboGrammar could point researchers in new directions, leading to more efficient and inventive designs.

RoboGrammar is described in a new study, and lead author Allan Zhao from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) is scheduled to present the software at the upcoming SIGGRAPH Asia conference. According to Zhao, robot design is still an overwhelmingly manual process, and people tend to fall back on the same conformations. “When you think of building a robot that needs to cross various terrains, you immediately jump to a quadruped,” says Zhao. RoboGrammar might have a different suggestion, though.

RoboGrammar runs through three steps before presenting its customized designs. To start, RoboGrammar needs a list of available parts and a task in the form of input terrains. For example, maybe you want to traverse terrain with ridges or steps. Next, the AI generates thousands of possible designs based on the available components. Most of these designs would be “nonsensical” robots that don’t work well with the specified terrain type (or much of anything). The team added a set of constraints called the “grammar graph” to ensure the designs created by RoboGrammar were functional on a basic level. Zhao says they took inspiration from animals, particularly arthropods, to focus the AI’s efforts.

Finally, RoboGrammar simulates all the designs with a controller algorithm called Model Predictive Control that prioritizes efficient forward movement. The researchers using RoboGrammar can search the database of possible designs with a “graph heuristic search” to find the best performers. They might have legs, wheels, or a mix of the two. Over time, the neural network learns which designs work well and which don’t, improving the heuristic function over time.

The designs that come from RoboGrammar aren’t finished products; they merely give engineers a better idea of which direction to go before they start building. Zhao also believes RoboGrammar could be useful in designing entirely virtual objects; with a different grammar graph it could just as easily churn out robots for a video game.

Continue reading

ET Deals: Dell Inspiron 15 5000 Intel Core i7-1165G7 Laptop for $674, iRobot Roomba i7+ 7550 Robot Vacuum for $599
ET Deals: Dell Inspiron 15 5000 Intel Core i7-1165G7 Laptop for $674, iRobot Roomba i7+ 7550 Robot Vacuum for $599

Today you can take advantage of a 10 percent discount to snag a Dell Inspiron 15 5000 laptop with an Intel Core i7-1165G7 processor, 12GB of RAM and a 512GB NVMe SSD for just $674. You can also get iRobot's Roomba i7+ robot vacuum for just $599.00, which is the same price it was on Cyber Monday.

Robot Dogs Now Fielding Robot Guns
Robot Dogs Now Fielding Robot Guns

It’s unclear what purpose they’re intended to fill—but the possibilities aren’t exactly pleasant.

Boston Dynamics Says Goodbye to 2020 With a Robot Dance Party
Boston Dynamics Says Goodbye to 2020 With a Robot Dance Party

One day, robots may be dancing on our graves, and they're going to be surprisingly good at it!

Hardware Accelerators May Dramatically Improve Robot Response Times
Hardware Accelerators May Dramatically Improve Robot Response Times

If we want to build better robots, we need them to be faster at planning their own motion. A new research team thinks it's invented a combined hardware/software deployment method that can cut existing latencies in half.