Researchers at MIT have unveiled a new system-on-a-chip that could revolutionize how tiny, low-power robots navigate complex environments. The chip, named Gleanmer, enables autonomous drones and other battery-limited devices to construct detailed 3D maps in real time using only about 6 milliwatts of power—roughly the energy of a single LED. This breakthrough, presented at the IEEE Very Large-Scale Integrated Circuits Symposium, promises to make small-scale robotics more capable and energy-efficient than ever.
A Smarter Approach to 3D Mapping
Traditional 3D mapping for robots relies on power-hungry systems that process and store thousands of 3D pixels (voxels) per image. The MIT team took a radically different approach by combining an efficient algorithm called GMMap with custom hardware optimized for its workload. Instead of using rigid cube-shaped voxels, GMMap represents obstacles and free space with flexible ellipsoidal blobs known as Gaussians. These shapes can stretch and shrink to match curved objects, storing far less data than traditional methods while still capturing accurate geometry.
“This paper showcases a key example of how you can leverage co-design of the algorithm and hardware to really push energy efficiency,” says Vivienne Sze, a professor in EECS and senior author of the paper. “Our chip allows you to store very large maps in a very small space, and do it in a very energy efficient manner.”
Hardware-Algorithm Co-Design at Its Best
The chip’s efficiency stems from several clever optimizations. First, the GMMap algorithm processes each depth image in a single pass, comparing only neighboring pixels rather than all pixels in an image. This drastically reduces the amount of data that needs to be stored and retrieved. Second, as the robot moves and observes the same object from different angles, overlapping Gaussians are fused into a single compact representation without re-examining the original pixels. Because Gaussians are more compact than raw pixel data, this fusion process runs on-chip with minimal memory traffic.
The hardware design keeps active Gaussians in small, fast on-chip memory right next to the processing units. “By having a dedicated memory that just stores the objects you’ve seen in the previous few frames, you can access the data much more efficiently,” explains co-lead author Zih-Sing Fu.
Unmatched Energy Efficiency
In tests, Gleanmer generated detailed 3D maps in real time while consuming only about 6 milliwatts of power—just 2.5 percent of what the best existing mapping chip requires. When used for path planning, the chip reuses compact Gaussians along the planned route, using only about 20 percent of the energy that conventional approaches would need.
The low power consumption makes Gleanmer ideal not only for tiny drones inspecting HVAC systems or pipelines, but also for lightweight augmented reality headsets that require continuous environmental mapping for educational medical simulations or repair work. Sertac Karaman, a professor of aeronautics and astronautics and co-author, notes: “Real-time 3D mapping has been the missing piece for small autonomous systems. Gleanmer makes that possible for the first time in a chip you can hold between your fingers.”
Looking Ahead
The researchers plan to push energy efficiency even further by moving processing units closer to the sensors that gather environmental data. They are also exploring applications beyond robotics, such as using Gaussians to represent complex schematics for AI reasoning systems. This work was supported by the MIT-MathWorks Fellowship, Amazon, the U.S. National Science Foundation, and Intel.
Leave a Reply