MIT Gleanmer chip: 6 mW SoC for tiny-robot 3D mapping

MIT researchers unveiled Gleanmer, a system-on-a-chip that constructs detailed 3D maps in real time while consuming about 6 milliwatts of power. The chip uses a compact Gaussian-based mapping algorithm to let battery-limited robots and lightweight AR headsets plan collision-free paths without storing whole images or large voxel grids.

How does Gleanmer create 3D maps?

Gleanmer represents obstacles using ellipsoid blobs called Gaussians rather than voxel cubes, and it generates those Gaussians in a single pass over depth images so the chip never has to store an entire image at once. The lab’s GMMap algorithm assumes nearby pixels belong to the same Gaussian, so the system only compares each pixel to its neighbors and can discard raw images after one pass. “At any point in time, we only need to store a few pixels in memory, which significantly reduces the memory footprint our algorithm requires,” says co-lead author Peter Zhi Xuan Li.

The researchers also fuse overlapping Gaussians directly, avoiding repeated access to raw pixels. That fusion keeps the map compact as the robot sees the same object from multiple viewpoints. Most computations operate directly on the compact Gaussians rather than on original pixels, which the team designed the hardware to accelerate.

How energy- and memory-efficient is Gleanmer?

Gleanmer consumes about 6 milliwatts of power, which the team says is only about 2.5 percent of the power required by the best existing chip for map construction. By reusing compact Gaussians along a planned path, the chip lets a robot chart a safe trajectory using about 20 percent of the energy it would otherwise need.

The hardware keeps the Gaussians it is actively working on in small, fast on-chip memory located beside the computational units, so data does not have to be fetched from power-hungry off-chip storage. That co-design of algorithm and hardware is central to the efficiency. Vivienne Sze, senior author on the paper, framed the approach as a demonstration of algorithm-hardware co-design that pushes energy efficiency while storing large maps in small space.

MIT’s team tested Gleanmer on diverse pre-existing 3D environments and also reconstructed obstacles and free space directly from live data streamed from an iPhone camera. The work was presented at the IEEE Very Large-Scale Integrated Circuits Symposium and appears in the paper titled "Gleanmer: A 6 mW SoC for Real-Time 3D Gaussian Occupancy Mapping."

Why it matters

Small autonomous systems have lacked a way to form detailed 3D maps continuously at very low power. Gleanmer targets that gap: its energy and memory profile makes real-time 3D mapping feasible for tiny UAVs inspecting confined industrial spaces, handheld AR headsets for extended use, or drones checking pipelines and HVAC systems for leaks. The chip’s single-pass, Gaussian-based mapping reduces the need for bulky memory and high energy budgets that have forced heavier compute platforms in the past.

What to watch

The researchers plan to push energy efficiency further by moving processing units closer to sensors and to explore new applications such as using Gaussians to represent schematics for efficient reasoning about blueprints. Watch for sensor-proximate SoC prototypes and demonstrations integrating Gleanmer into drones or AR headsets as the next milestones.