NeuSLAM: Dense Visual SLAM on Edge Devices

Aniket Gupta, Tianye Ding, Ajay Rajendra Kumar, Aruni RoyChowdhury, Charles Saunders, Hanumant Singh, Huaizu Jiang, Pia K Bideau, Qu Cao

Keywords: SLAM, Edge Devices, Jetson

TL;DR: We present a Dense Visual SLAM method which runs at real-time (>10 FPS) on Jetson Orin Nano and blazing speeds (>120 FPS) on desktops.

Abstract: We present NeuSLAM, a hybrid architecture for dense visual SLAM designed specifically for stereo and RGB-D sensors on resource-constrained edge devices. While recent learning-based dense SLAM methods achieve strong trajectory accuracy and rich scene reconstruction, their learned back-ends typically maintain dense correlation volumes and feature maps on the GPU throughout optimization. This requires several gigabytes of memory, restricting them to desktop-grade hardware. To overcome this bottleneck, we introduce a lightweight neural network extending NeuFlow-V2, that jointly predicts dense optical flow and stereo disparity alongside per-pixel confidence maps via a shared feature encoder. For RGB-D setups, the disparity branch is simply bypassed in favor of sensor depth. Our front-end isolates high-confidence sparse correspondences for camera pose estimation, while a classical back-end maintains a lightweight pose graph over keyframes to ensure global consistency and keep GPU memory usage minimal. In zero-shot synthetic-to-real evaluations, our network achieves a state-of-the-art F1-all score of 9.9\% for optical flow on KITTI 2015, alongside competitive stereo disparity on both KITTI 2015 and ETH3D. On TUM RGB-D and FLSea-VI datasets, NeuSLAM achieves trajectory accuracy competitive with DROID-SLAM while running 6.5$\times$ faster and consuming 87\% less GPU memory. Our model and code will be open sourced.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Paper Acceptance: Yes

Submission Number: 25