3D occupancy maps are essential for autonomous navigation and augmented reality (AR) applications, as they classify environments into occupied, unexplored, and free regions. Unlike sparse maps that contain only a few points within occupied areas, occupancy maps capture all spatial regions, enabling users and their devices to navigate carefully while staying within navigable regions and avoiding unexplored or occupied ones.
Traditional approaches such as OctoMap use voxels (3D pixels) to represent space, which consume significant memory and power. Even with hardware acceleration, these methods exceed 100 mW, far surpassing typical edge power budgets of under 10mW.
GMMap offers a more compact representation using Gaussian ellipsoids, where occupied regions are represented by red Gaussians while free regions use blue Gaussians. These ellipsoids, characterized by their covariances and means, flexibly adapt to object geometry, making them far more efficient than rigid cube-shaped voxels.
A single elongated Gaussian can represent a region that would otherwise require many voxels. However, software implementations of GMMap on embedded systems consume more than 2 W, making them prohibitive for battery-limited devices. This power consumption gap motivated the development of Gleanmer as a dedicated hardware solution.
Algorithm-Hardware Co-Design for Energy Efficiency
The first major innovation addresses the energy-intensive generation of free Gaussian bases. Traditional GMMap implementations generate these bases from sensor rays, which are affiliated with line segments extracted from depth images.
This computational cost increases with the number of line segments and accounts for around 26% to 65% of the total energy used for map construction. Instead of storing and processing sensor rays, Gleanmer's free Gaussian bases generation unit computes free Gaussian bases from occupied Gaussians by sampling a small pool of representative rays from each occupied Gaussian.
This technique avoids the memory overhead of storing thousands of line segments, yielding energy savings of 22% to 63% during map construction. By operating directly on the compact Gaussian representation rather than raw pixel data, the chip achieves significant energy reductions without compromising accuracy.
The second major innovation accelerates map queries through batch processing. After construction, devices query occupancy probabilities at coordinates along planned trajectories to ensure collision-free motion. These spatially proximate coordinates typically navigate similar R-Tree paths when retrieving overlapping Gaussians.
Gleanmer implements a batch-querying strategy in which a single bounding box enclosing 16 consecutive coordinates retrieves all overlapping Gaussians simultaneously via an R-Tree. Throughout all coordinates, these Gaussians are time-interleaved, sharing the Gaussian Regression Unit with only a 2% area overhead for registers. This batching increases query throughput by four to 10 times while reducing energy consumption.
Approximate Computing and Silicon Implementation
The Gleanmer architecture leverages approximate computation to reduce silicon area while maintaining map accuracy. The buffer area in Gaussian generation is reduced eightfold, thanks to two innovations.
First, producing free Gaussian bases directly from occupied Gaussians eliminates storing bases for line segments. Second, a single-cycle slope approximation in the scanline segmentation (SS) unit allows line segments from a single depth-image row to be buffered, rather than the four rows traditionally required.
The SS unit typically needs four cycles to compute the slope, necessitating time interleaving across four rows. By approximating the slope in a single cycle using estimates from the previous four cycles, the SS unit achieves equivalent throughput without time-interleaving. Decreasing Gaussian precision (32-bit to 19-bit) produces around a 38% reduction in the fusion and regression engine area, while covariance matrices remain 32-bit to prevent degeneracy.
Together, these approximations reduce the accelerator area by 38% and shrink the map size by around 44% to 63%, while maintaining accuracy across diverse environments. Implementation results demonstrate exceptional performance.
Fabricated in 16nm CMOS, the chip processes 640×480 depth images beyond 88 fps during construction and handles over 540,000 coordinates per second (cps) during queries, reaching up to 1.32 million cps. Testing across three environments, including live data from an iPhone camera, validates the chip's ability to generate detailed occupancy maps with accuracy comparable to existing implementations.
From Research to Real-World Deployment
Gleanmer represents a significant advancement in enabling real-time 3D occupancy mapping for resource-constrained edge devices. By combining an algorithm that operates directly on compact Gaussian representations with specialized hardware designed to accelerate that specific workload, the researchers have demonstrated that high-fidelity mapping need not require hundreds of milliwatts of power.
The chip's ability to construct maps at more than 88 fps while consuming only 6mW makes it feasible for tiny drones to navigate complex environments. The success of this approach underscores the importance of algorithm-hardware co-design in pushing energy efficiency boundaries.
The research team plans to further improve efficiency by moving processing units closer to sensors and exploring additional applications like schematic representation for AI systems. The paper demonstrates that with thoughtful co-optimization, the computational demands of mapping need not stand in the way of deploying sophisticated autonomous capabilities on the smallest of devices.
Journal Reference
Fu, Z.-S., Peter, L., Karaman, S., and Sze, V. (2026). Gleanmer: A 6 mW SoC for Real-Time 3D Gaussian Occupancy Mapping. ArXiv.org. https://arxiv.org/abs/2603.29005.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.