MIT Develops 6-Milliwatt Ultra-Low-Power Chip That Enables Real-Time 3D Mapping on Micro-Drones

Researchers at the Massachusetts Institute of Technology (MIT) have developed an ultra-low-power chip that consumes just 6 milliwatts of electricity, enabling micro-drones and robots to build detailed 3D maps of their surroundings in real time. The system-on-chip (SoC), named Gleanmer, could allow battery-powered autonomous machines to navigate cluttered environments with ease — including industrial ventilation systems, warehouses, tunnels, and other confined spaces where precise obstacle avoidance is critical.

The technology also holds promise for lightweight augmented reality (AR) headsets, enabling real-time indoor mapping without significantly draining battery life. The chip combines dedicated hardware with a compact mapping algorithm to dramatically reduce the memory and power required for robots to construct 3D environmental models.

Breaking Through Traditional Mapping Bottlenecks

Building detailed 3D maps typically requires a robot to process large volumes of image data and store complex environmental models — a memory and power burden that small, battery-powered devices often cannot bear.

Rather than using the conventional approach of representing environments with millions of tiny cubes (voxels), the MIT team adopted flexible ellipsoidal shapes called Gaussians. These shapes represent curved objects and open spaces more efficiently than voxels while requiring far less memory.

The researchers paired the chip with a mapping algorithm called GMMap, which constructs 3D maps from depth images in a single pass. This allows the system to discard raw image data almost immediately after capture, eliminating the need for repeated storage and processing.

"At any point in time, we only need to store a small number of pixels in memory, which dramatically reduces the memory footprint required by the algorithm," said Peter Zhi Xuan Li, co-first author of the study.

The system also addresses another common challenge in mapping: as a robot moves, it often observes the same object from multiple angles, generating overlapping model representations that inflate map size. The MIT team developed a method to merge overlapping Gaussians directly, without referencing the original image data.

Small Chip, Big Impact

This design allows researchers to keep most active data in fast on-chip memory rather than relying on power-hungry external storage. "By using a dedicated memory that only stores objects observed over the past few frames, data can be accessed much more efficiently," said co-first author Zih-Sing Fu.

In tests conducted across a variety of pre-recorded environments, Gleanmer generated detailed 3D maps in real time at approximately 6 milliwatts — roughly 2.5% of the energy consumed by the best comparable mapping chips currently available, according to the researchers.

The chip is also capable of reconstructing obstacles and free space directly from live iPhone camera streams. By reusing the compact Gaussian representation during path planning, the system enables robots to calculate collision-free routes using approximately 20% of the energy that would normally be required.

"This paper demonstrates a key example of how algorithm and hardware co-design can truly and substantially improve energy efficiency," said Vivienne Sze, MIT professor of Electrical Engineering and Computer Science and senior author of the study.

The research team believes future versions could achieve further efficiency gains by moving computation closer to onboard sensors. Beyond robotics, the team is also exploring whether Gaussian representations could help computer systems process technical drawings and complex circuit diagrams more efficiently.

The research was officially presented at the IEEE Very Large-Scale Integrated Circuits (VLSI) Symposium.

原文來源： 查看原文