New Low Power Chip Helps Robots Navigate Tricky Environments

Accurate 3D maps are essential for many AR, VR, autonomous navigation, and machine vision applications. However, many mapping systems use significant energy and require a lot of memory to build and store the 3D maps. Researchers at MIT have developed a new system on chip (SoC) called Gleanmer, that uses a GMMap accelerate that can significantly reduce power and memory needs.

By combining advanced algorithms with dedicated hardware, the team generated 3D maps that could help robots (small UAVs) navigate while using minimal memory and power―a reduction in construction energy by up 63% and query energy by 81%.

The new chip allows small autonomous robots to navigate complex environments with a lot of tight corners and obstacles, using about the same amount of power as an LED. The chip constructs detailed real-time 3D maps of their local environment allowing UAVs to navigate tricky environments with more confidence, for example, inside an industrial HVAC system to look for gas leaks.

This low energy and memory approach was made possible by combining the efficient GMMap mapping algorithm― developed in the same laboratory―with hardware designed to accelerate its workload. As a result, it only consumes around 6 mW of power while still being able to store the maps. The map is also not represented using cubic 3D pixels, instead using ellipsoid blobs called Gaussians that match the geometry of curved objects more efficiently than cubic pixels. 

The map captures both the obstacles and free space around the robot, making it easy for the robot to chart a collision-free path. The images are captured in one pass (unlike pixel-based maps), and the images then discarded, eliminating high memory requirements. Instead of comparing all the voxel (cubic) pixels with each other in the image, the Gaussian ellipsoid pixels are only compared with their neighbors, so only a few pixels need to be stored in the memory at any one time.

ROI Calculator

Discover the potential cost savings of robotic automation over a 20-year system life

This calculator compares your current manual labor costs against the total cost of owning and operating a robotic system over its 20-year lifespan.

EXPLORE TODAY

When the robot moves through a space, it sees an object from different viewpoints, and the Gaussians overlap and fuse, making the map more compact and better suited for edge devices. This also contributes to reduced memory and power requirements, alongside discarding pixels. Gaussians are also smaller than cubic pixels and is another reason why the power and memory requirements are lower. This also generated a more compact map size than cubic pixel maps, with the map sizes shrinking by 44–63% while maintaining map accuracy. It is this combination of factors that enabled the chip to be suitable for small autonomous robots with limited power and memory storage capabilities, as all the data can be stored using the on-chip memory without the need for extra memory storage.

The real-time 3D mapping makes it a lot easier for autonomous systems to navigate complex environments. The technology could also be applied to AR/VR glasses to help wearers better navigate the space around them in the virtual environment. Even though the robotic ‘vision’ system is already very energy efficient with low memory requirements, there are plans to further improve it. This includes moving the processing units closer to the sensors (that gather the local environmental data) on the chip. There are also plans to expand the application use cases of the chip, venturing into using Gaussians for representing schematics, which could help AI systems to be more efficient at analyzing complex blueprints.