A research team at MIT has unveiled an ultra-low-power chip that lets small robots and drones construct detailed 3D maps of their surroundings in real time using barely more electricity than a standard LED. The innovation could help autonomous flying robots inspect tight industrial spaces such as HVAC ducts for gas leaks, plan collision-free routes, and operate for extended periods without recharging.
Unlike conventional systems that rely on power-hungry processors and large amounts of memory to build and store dense 3D maps using rigid cubes called voxels, MIT’s approach combines a compact mapping algorithm with specialized hardware to slash energy and memory demands.
At the heart of the solution is a system-on-a-chip named Gleanmer that consumes just 6 milliwatts—roughly one-thousandth the power of comparable platforms. The chip’s efficiency makes it a strong candidate for lightweight augmented reality headsets used in medical training or equipment repair, where long battery life is critical.
“This work demonstrates how co-designing algorithms and hardware can push energy efficiency to new limits,” says Vivienne Sze, a professor in MIT’s Department of Electrical Engineering and Computer Science, a member of the Research Laboratory of Electronics, and the paper’s senior author. “Our chip stores large maps in a tiny footprint while keeping energy use minimal—something previous methods struggled to achieve.”
Sze is joined by co-lead authors and MIT graduate students Zih-Sing Fu and Peter Zhi Xuan Li, along with Sertac Karaman, professor of aeronautics and astronautics and director of the Laboratory for Information and Decision Systems. The findings were presented at the IEEE Very Large-Scale Integrated Circuits Symposium.
Mapping with flexible ellipsoids instead of rigid cubes
Traditional 3D mapping techniques represent obstacles as stacks of rigid cube-shaped voxels. Each voxel must be processed multiple times, consuming significant power and memory—especially when the robot revisits an area from a different angle.
The MIT team replaces voxels with flexible ellipsoid blobs called Gaussians. These shapes can stretch, bend, and twist to match curved surfaces more naturally than rigid cubes, reducing the number of data points needed to describe an object. A single elongated ellipsoid can stand in for many voxels, sharply cutting both memory usage and computational load.
The system’s core algorithm, GMMap, builds these Gaussians from depth camera images in a single pass, then discards the raw frames entirely. Instead of comparing every pixel to every other pixel—an expensive operation—it groups nearby pixels into the same Gaussian, comparing each pixel only to its immediate neighbors.
“By storing only a handful of pixels at any moment, we keep the memory footprint tiny,” says Li. “This lets the algorithm run efficiently even on devices with very limited resources.”
Keeping maps compact as robots move
As a drone explores, it often sees the same object from different angles. In voxel-based systems, overlapping data accumulates quickly, bloating memory requirements. The MIT team solves this by fusing overlapping Gaussians directly, without reloading the original pixel data.
Because Gaussians are far more compact than raw pixels, the fusion process runs faster and uses less power. The chip stores only the most recent Gaussians in fast, on-chip memory located next to the processing units—eliminating the need to fetch data from slower, energy-intensive off-chip storage.
“Having a dedicated memory buffer for the objects you’ve seen in the last few frames lets you access data much faster and with far less energy,” explains Fu.
Validated across diverse 3D environments
The researchers evaluated Gleanmer by reconstructing a variety of pre-existing 3D indoor and industrial environments. The chip consistently generated accurate maps while consuming minimal power, outperforming state-of-the-art alternatives in both energy efficiency and memory footprint.
Looking ahead, the team sees applications beyond robotics—from portable medical simulators to wearable AR systems that guide technicians through complex repairs. By rethinking how maps are built and stored, the new chip paves the way for smarter, longer-lasting devices that can operate where power is scarce and precision is essential.
AI summary
MIT araştırmacıları, 6 milivat güçle çalışan yeni bir çip sayesinde minyatür robotların karmaşık ortamlarda güvenle dolaşmasını ve 3D haritalar oluşturmasını sağladı. Bu yenilik, otonom cihazların enerji verimliliğini artırırken, artırılmış gerçeklik gözlükleri gibi uzun süreli kullanım gerektiren uygulamalarda da devrim yaratabilir.
