Small autonomous robots need to understand the world around them quickly, but detailed 3D mapping has usually demanded more power and memory than tiny, battery-limited devices can spare. MIT researchers have developed a chip aimed at changing that tradeoff.
The system-on-a-chip, called Gleanmer, builds detailed 3D maps in real time while using only about as much power as a single LED. In practical terms, that could help a tiny, low-power UAV move through tight areas such as an industrial HVAC system while avoiding obstacles and checking for gas leaks.
Why 3D mapping is hard for small robots
For a robot to move safely, it needs more than a rough picture of its surroundings. It needs to know where obstacles are, where free space is, and how to plan a collision-free path toward its goal.
Traditional 3D mapping methods often represent space with voxels, which are cube-shaped 3D pixels. That approach can be accurate, but it can also be expensive for small devices because the robot must store camera images and process many 3D pixels multiple times.
That memory burden matters at the edge. When a device is small, lightweight, and battery-limited, power-hungry map construction can make real-time autonomy difficult.
How Gleanmer makes maps compact
The MIT team took a different route by combining an efficient mapping algorithm with hardware built specifically to accelerate it. Instead of using voxels to represent the environment, the system maps obstacles with ellipsoid blobs called Gaussians.
Gaussians can change size, shape, and thickness, which lets them fit curved objects more efficiently than rigid cubes. A single elongated ellipsoid can represent a region that would otherwise require many voxels.
This matters because the map needs to capture both obstacles and free space. Together, those elements allow a robot to plan a safe route. By representing occupied surfaces and free space more compactly, the system reduces the memory needed to keep the map useful.
Gleanmer uses GMMap, an algorithm developed in the researchers' lab, to generate 3D maps with Gaussians. The chip does not need to store an entire depth image at once. Instead, the algorithm can create accurate Gaussians from depth images in one pass and then discard the images.
A key assumption makes that possible: nearby pixels are treated as likely to belong in the same Gaussian. That means each pixel only needs to be compared with its neighbors rather than with every other pixel in the 3D image.
Co-design is the point
The advance is not just the map format or just the chip. The researchers designed the algorithm and hardware together so the workload fits the physical chip efficiently.
As a robot moves, it may see the same object from multiple viewpoints. That can create overlapping Gaussians that describe the same object, making the map larger than necessary. The team developed a method to fuse overlapping Gaussians directly, without going back to the original raw pixels.
Because the chip works mostly with compact Gaussians rather than full image data, it can keep the active data close to the computation units in small, fast on-chip memory. That avoids repeated fetches from more distant, power-hungry off-chip storage.
The result is a system designed around memory efficiency from the start. The algorithm reduces what must be stored, and the hardware accelerates the work that remains.
What the chip achieved
The researchers tested Gleanmer by reconstructing a range of diverse, pre-existing 3D environments. The chip can also reconstruct obstacles and free space from live data streamed from an iPhone camera.
In those tests, Gleanmer generated detailed 3D maps in real time while consuming about 6 milliwatts of power. The source article says that was only about 2.5 percent of the power required by the best existing chip for map construction.
The energy savings also extend to planning. By reusing compact Gaussians along a path, the chip lets a robot chart a safe trajectory using only about 20 percent of the energy it would otherwise need.
The work was recently presented at the IEEE Very Large-Scale Integrated Circuits Symposium. The paper's senior author is Vivienne Sze, a professor in the Department of Electrical Engineering and Computer Science (EECS) and a member of the Research Laboratory of Electronics (RLE). Co-lead authors are MIT graduate students Zih-Sing Fu and Peter Zhi Xuan Li, joined by Sertac Karaman, a professor of aeronautics and astronautics and the director of LIDS.
Where it could go next
The most direct use case is small autonomous robotics. A tiny UAV moving through a tight industrial space needs to understand its surroundings continuously without carrying heavy computing hardware or draining its battery quickly.
The same low-power mapping capability could also fit lightweight augmented reality headsets. The source article points to applications such as educational medical simulation and detailed repair and assembly work, where a headset may need to be worn for extended periods.
The researchers plan to improve energy efficiency further by moving the processing units on the chip closer to the sensors that gather environmental data. They may also explore using Gaussians to represent schematics, which could help AI systems reason about complex blueprints more efficiently.
For small robots and other edge devices, the larger implication is straightforward: real-time spatial understanding becomes more practical when the map itself is compact and the hardware is built around that compact form.