MIT researchers have developed a new system-on-a-chip called Gleanmer that enables small, battery-powered robots and devices to generate detailed 3D maps of their surroundings in real-time. By combining specialized hardware with an efficient mapping algorithm, the chip achieves this high-performance navigation while consuming only about 6 milliwatts of power. This energy efficiency allows it to operate using a fraction of the power required by existing mapping systems, making it suitable for lightweight autonomous drones and wearable augmented reality headsets.

Conventional 3D mapping typically demands significant memory and power, as devices must store and repeatedly process high-resolution depth images captured by cameras. To overcome these limitations, the MIT team utilized a representation technique based on Gaussians (ellipsoid-shaped blobs) rather than traditional rigid voxels (3D cube pixels). Because these flexible ellipsoids can adapt to curved surfaces and object geometry more effectively, the system requires significantly less storage. The team's proprietary algorithm, GMMap, processes depth images in a single pass by comparing each pixel only to its immediate neighbors, eliminating the need to store entire images in memory.

The chip architecture further optimizes energy usage through a co-design approach where memory is positioned directly alongside computational units. This proximity ensures that frequently accessed Gaussian data remains on-chip, preventing the power-intensive process of fetching information from external storage. During testing in diverse 3D environments and with live data from an iPhone camera, the system demonstrated significant efficiency gains. Gleanmer required only about 2.5 percent of the power used by the best existing mapping chips. Furthermore, by reusing compact Gaussian data for path planning, the chip allows robots to chart safe trajectories using approximately 20 percent of the energy typically consumed by traditional methods.

The project, which was presented at the IEEE Very Large-Scale Integrated Circuits Symposium, aims to bring continuous, real-time spatial awareness to small-scale hardware. The researchers are now exploring ways to further enhance efficiency by integrating processing units closer to environmental sensors. They also intend to investigate whether Gaussians can be used to represent schematics, potentially assisting AI systems in reasoning about complex blueprints more effectively.

MIT researchers have developed a new system-on-a-chip called Gleanmer that enables small, battery-powered robots and devices to generate detailed 3D maps of their surroundings in real-time. By combining specialized hardware with an efficient mapping algorithm, the chip achieves this high-performance navigation while consuming only about 6 milliwatts of power. This energy efficiency allows it to operate using a fraction of the power required by existing mapping systems, making it suitable for lightweight autonomous drones and wearable augmented reality headsets.

Conventional 3D mapping typically demands significant memory and power, as devices must store and repeatedly process high-resolution depth images captured by cameras. To overcome these limitations, the MIT team utilized a representation technique based on Gaussians (ellipsoid-shaped blobs) rather than traditional rigid voxels (3D cube pixels). Because these flexible ellipsoids can adapt to curved surfaces and object geometry more effectively, the system requires significantly less storage. The team's proprietary algorithm, GMMap, processes depth images in a single pass by comparing each pixel only to its immediate neighbors, eliminating the need to store entire images in memory.

The chip architecture further optimizes energy usage through a co-design approach where memory is positioned directly alongside computational units. This proximity ensures that frequently accessed Gaussian data remains on-chip, preventing the power-intensive process of fetching information from external storage. During testing in diverse 3D environments and with live data from an iPhone camera, the system demonstrated significant efficiency gains. Gleanmer required only about 2.5 percent of the power used by the best existing mapping chips. Furthermore, by reusing compact Gaussian data for path planning, the chip allows robots to chart safe trajectories using approximately 20 percent of the energy typically consumed by traditional methods.

The project, which was presented at the IEEE Very Large-Scale Integrated Circuits Symposium, aims to bring continuous, real-time spatial awareness to small-scale hardware. The researchers are now exploring ways to further enhance efficiency by integrating processing units closer to environmental sensors. They also intend to investigate whether Gaussians can be used to represent schematics, potentially assisting AI systems in reasoning about complex blueprints more effectively.