MIT's Gleanmer Chip Enables Real-Time 3D Mapping for Tiny Robots

MIT researchers have developed a new system-on-a-chip, known as Gleanmer, that enables tiny, battery-limited robots to construct detailed 3D maps of their surroundings in real-time. By combining a highly efficient mapping algorithm with specialized hardware, the chip performs complex navigation tasks while consuming only about 6 milliwatts of power—a fraction of the energy required by existing systems. This breakthrough could allow small autonomous devices to safely navigate complex environments, such as industrial HVAC systems, or support lightweight augmented reality headsets.

Rethinking 3D Mapping Efficiency

Traditional 3D mapping typically requires significant memory and power because robots must store and repeatedly process high-resolution images of their environment. These systems often represent obstacles using voxels, which are rigid, cube-shaped pixels that demand substantial computational resources. To overcome these limitations, the MIT team utilized a technique that represents obstacles as ellipsoid blobs called Gaussians.
Because these ellipsoids can be smoothly adapted to match the shape of curved objects, they are far more compact than voxels. A single elongated Gaussian can represent a region that would otherwise require many voxels to define. By mapping both obstacles and free space with these flexible shapes, the system significantly reduces the memory footprint required to plan a collision-free path.

Streamlined Processing and Hardware Co-Design

The Gleanmer system relies on an algorithm called GMMap, which generates accurate Gaussians from depth images in a single pass. Unlike traditional methods that compare every pixel in an image to every other pixel, this algorithm assumes that nearby pixels belong to the same Gaussian. By only comparing pixels to their immediate neighbors, the chip avoids the need to store entire images in memory at once.
To further enhance efficiency, the researchers designed the hardware to perform data fusion directly on the Gaussians themselves, rather than revisiting the original raw pixels. By keeping these compact Gaussians in fast, on-chip memory located near the computational units, the system avoids the power-intensive process of fetching data from distant, off-chip storage.

Real-World Performance and Future Potential

In testing, the Gleanmer chip successfully reconstructed diverse 3D environments and processed live data streamed from an iPhone camera. The system achieved real-time mapping while using only about 2.5 percent of the power required by the best existing chips for similar tasks. Furthermore, by reusing these compact Gaussians during path planning, the chip allows robots to chart safe trajectories using only 20 percent of the energy typically needed.
The research team, led by Professor Vivienne Sze and Professor Sertac Karaman, suggests that this technology could eventually be applied to help AI systems reason about complex blueprints or schematics. By moving processing units closer to environmental sensors, the researchers aim to push energy efficiency even further, potentially enabling a new generation of autonomous devices that can understand their surroundings continuously and at minimal power cost.