MIT researchers have developed a chip that lets miniature robots and other low-power devices construct detailed 3D maps of their surroundings in real time while drawing roughly 6 milliwatts, about the energy of a single LED. Presented at the IEEE Symposium on VLSI Technology and Circuits, the chip, called Gleanmer, could open constrained spaces like ventilation shafts, mineshafts, and warehouse racking to tiny autonomous machines.
Ellipsoids Instead of Voxels
Conventional 3D mapping pipelines carve space into voxels, an approach whose memory and compute appetite quickly outgrows what a palm-sized robot can carry. Gleanmer instead represents obstacles and free space with Gaussian ellipsoids, compact statistical shapes that capture the geometry of a cluttered environment with a fraction of the data. The researchers co-designed an efficient mapping algorithm with dedicated hardware, allowing the integrated chip to generate navigation-ready maps using minimal memory and power.
Why It Matters for Micro-Robotics
Depth cameras and lidar have shrunk fast, but the computation to fuse them into maps has remained a power hog, forcing designers of insect-scale and hand-sized robots to choose between autonomy and endurance. A 6-milliwatt mapping budget changes that calculus: the chip could help tiny robots avoid obstacles while inspecting industrial HVAC systems for gas leaks, or navigate collapsed structures where batteries must last for hours. The work slots into a broader push toward capable perception at the edge, seen in Efinix's Titanium Edge FPGAs for constrained AI and Luxonis's OAK cameras for physical AI.
From Lab Demo to Deployment
The team demonstrated real-time mapping performance on the fabricated chip rather than in simulation, an important step toward integration into commercial micro-drones and inspection robots. Cheap, ultra-efficient spatial awareness is a missing primitive for the swarms of small machines expected to handle inspection and monitoring tasks, and it complements the data-side efforts fueling robot intelligence, such as X Square Robot's 2,000-hour open-source manipulation dataset.
Reporting based on coverage from MIT News and TechXplore.