Researchers at the Massachusetts Institute of Technology (MIT) have developed an ultra-efficient computer chip that could help tiny robots and lightweight devices create detailed 3D maps of their surroundings while using very little energy.
The new chip could allow small drones to fly through narrow spaces, such as air-conditioning ducts or industrial pipelines, to inspect equipment and look for problems like gas leaks.
It may also improve lightweight augmented reality (AR) glasses and other battery-powered devices.
One of the biggest challenges for small robots is understanding their surroundings.
To move safely, a robot needs to build a three-dimensional map that shows obstacles and empty spaces. It can then use this map to plan a path and avoid collisions.
Creating these maps usually requires a lot of computing power and memory. Traditional systems process huge amounts of image data and store large numbers of tiny cubes called voxels, which are like 3D pixels.
Because so much information must be stored and processed, these systems consume significant amounts of energy.
MIT researchers took a different approach.
Instead of building maps using countless tiny cubes, they developed a method that uses flexible, blob-like shapes called Gaussians. These shapes are similar to stretched or compressed ellipsoids and can adapt to the curves and contours of real objects.
Because one Gaussian can represent an area that would normally require many voxels, the maps become much smaller and more efficient. At the same time, they still accurately show obstacles and open spaces that robots need to navigate safely.
The researchers designed a special chip called Gleanmer to take advantage of this mapping method. The chip uses an algorithm called GMMap, which can create highly detailed 3D maps in real time.
The algorithm also avoids another major problem. Traditional systems often need to examine the same images repeatedly and store large amounts of data in memory. The new method only needs to process each depth image once. After the necessary information has been extracted, the image can be discarded.
Instead of comparing every pixel in an image with every other pixel, the system assumes that nearby pixels usually belong to the same object. This dramatically reduces the amount of computing required.
As a robot moves, it often sees the same object from different angles. Normally, this would create duplicate information and make maps larger. The MIT team developed a way to combine overlapping Gaussians directly, without having to return to the original image data. This further reduces memory requirements and energy use.
The chip keeps the information it is actively using in small, fast memory units located directly next to its computing components. This allows it to access data quickly without repeatedly retrieving information from larger, power-hungry storage systems.
The results are impressive. Gleanmer consumes only about six milliwatts of power, roughly the same amount of electricity used by a single LED light. It uses only about 2.5% of the energy required by the best existing map-building chips.
The chip has already demonstrated its ability to create detailed maps from live camera data, including video streamed from an iPhone.
The researchers believe this technology could open new possibilities for small autonomous systems.
Tiny inspection drones, wearable AR devices, and other battery-powered technologies may soon be able to understand and navigate complex environments in real time while consuming remarkably little energy.
