MIT researchers combined a custom algorithm with silicon built to accelerate it, letting insect sized drones and lightweight AR headsets build dense 3D maps of tight spaces on a milliwatt scale power budget.
A new chip from MIT draws less power than a single indicator LED, and in the process it does something that power-constrained devices have never managed well: it builds a full three-dimensional map of its surroundings in real time, dense enough for a tiny robot to plan an entire collision-free route through a space as cramped as an air duct.
The system-on-a-chip consumes about 6 milliwatts, roughly the draw of one LED on a household device. The research team at MIT paired a highly efficient 3D-mapping algorithm with custom hardware that accelerates that algorithm, shrinking both the memory and the energy normally needed to store and process a robot's view of obstacles. The result is not a small efficiency bump. It is the simultaneous breaking of two constraints that have always traded off against each other in milliwatt-class systems.
According to MIT's announcement, the insight sits at the silicon layer. By designing the algorithm and the hardware together, the team can build a 3D map dense enough to plan an entire collision-free path, not just dodge the nearest obstacle, while staying inside a power budget that fits on an insect-sized device.
The team's most concrete target is small autonomous UAVs that can navigate tight, obstacle-rich spaces such as industrial HVAC ductwork to find gas leaks. Today, inspections of that kind are done by people in protective gear or by larger drones that cannot fit into the duct. A milliwatt-scale mapper is what makes an insect-class inspector physically possible, because the robot has to carry its own power and its own intelligence.
The same silicon also fits inside lightweight augmented-reality headsets meant to be worn for long stretches. The researchers point to educational medical simulation, repair work, and assembly-line training as target applications where current AR hardware is too heavy and too power-hungry to wear for a full shift.
The work comes from the group of Vivienne Sze, a professor in MIT's Department of Electrical Engineering and Computer Science and a member of the Research Laboratory of Electronics. The co-lead authors are Zih-Sing Fu and Peter Zhi Xuan Li. Sertac Karaman, who co-directs the Laboratory for Information and Decision Systems and works on autonomous aerial vehicles, is also a co-author. The team will present the chip at the IEEE VLSI Symposium.
The contribution is best understood as a structural one, not a benchmark win. Existing low-power mapping systems either drop to a thin slice of 3D, enough to dodge the nearest wall, or burn power that rules out always-on, size-constrained devices. This chip keeps the dense map and stays milliwatt-class by moving the heavy lifting into dedicated hardware that knows exactly how the algorithm will use it. That is a different shape of problem than throwing a better general-purpose processor at the task.
The limits are worth naming. The 6-milliwatt figure and the map density come from MIT's own announcement and a forthcoming conference paper. Independent benchmarks, third-party replication, and any commercial timeline are not part of the public source basis. Gas-leak inspection and long-wear AR are stated targets, not deployed products. The chip shows what co-design can do in a controlled setting. Whether it survives real-world temperature swings, dust, vibration, and the certification pathways for industrial or medical use is an open question.
What to watch next: the IEEE VLSI paper itself, when it is available, and any partnership with HVAC robotics or AR hardware vendors willing to put the silicon in a real device. Either step would move this from a research milestone to a deployment conversation.