MIT researchers have created a new chip that could assist small, low-power UAVs in navigating obstacles as they maneuver through complex environments, such as industrial HVAC systems, to inspect for gas leaks. This chip enables tiny autonomous robots and other battery-dependent devices to generate detailed 3D maps of their surroundings in real-time, consuming only as much power as an LED. Such maps allow robots to chart collision-free routes to achieve their objectives.<\/p>\n
Traditionally, constructing these comprehensive maps requires systems that consume significant power and memory to develop and store 3D representations of obstacles. The MIT team adopted a novel strategy by integrating an efficient mapping algorithm with specialized hardware to expedite its processes, thereby reducing memory and power usage. The resulting system-on-a-chip uses merely 6 milliwatts of power, significantly less than other systems, making it suitable for applications like lightweight augmented reality headsets used in educational medical simulations or intricate repair work.<\/p>\n
Vivienne Sze, a professor in the EECS Department and RLE member, along with co-lead authors Zih-Sing Fu and Peter Zhi Xuan Li, and Sertac Karaman, a professor of aeronautics and astronautics, presented this work at the IEEE VLSI Symposium. The chip, known as Gleanmer, utilizes the GMMap algorithm to efficiently create a 3D map of a robot\u2019s environment using Gaussians to represent obstacles. This method circumvents the need to process each depth image multiple times, allowing for memory-efficient operation.<\/p>\n
Instead of employing 3D pixels, the researchers mapped obstacles using ellipsoid shapes called Gaussians, which adapt to the contours of objects more effectively than traditional voxel-based methods. This approach allows for a more compact representation of the environment, reducing memory demands and power consumption significantly. The algorithm assumes neighboring pixels belong to the same Gaussian, streamlining the process by only comparing each pixel to its neighbors.<\/p>\n
As robots move and view objects from various angles, overlapping Gaussians can make maps too large for edge devices. The team developed a technique to merge these Gaussians without revisiting original pixels, further minimizing memory and power needs. This efficiency is maintained throughout the algorithm, with computations focused on compact Gaussians instead of the original pixels.<\/p>\n
The chip stores active Gaussians in small, fast on-chip memory near computational units, eliminating the need for power-intensive off-chip storage. Fu explains that this dedicated memory efficiently accesses data from recent frames. Testing showed the chip could reconstruct 3D environments from live data with iPhone cameras, using only 2.5% of the power required by the best existing mapping chip.<\/p>\n
Li notes that the chip reduces energy use by ensuring algorithm efficiency and accelerating the workload. Future improvements could involve moving processing units closer to sensors and exploring Gaussians for representing schematics, aiding AI in reasoning about blueprints. Karaman highlights the chip’s potential for real-time 3D mapping in small autonomous systems, supported by MIT-MathWorks Fellowship, Amazon, NSF, and Intel.<\/p>\n