NE-VSLAM Edge Acceleration Platform

Real-time Visual SLAM on resource-constrained edge devices. We accelerate Visual SLAM through two complementary methods: software optimization for specific CPU architectures, and selective FPGA-based hardware acceleration. This lets robots localize and map where GPS and human control are unavailable.

RGB-D tracking on a Raspberry Pi 4
31 FPS

RGB-D tracking on a Raspberry Pi 4

Faster than the tested baselines
2.1 ×

Faster than the tested baselines

Trajectory error (ATE RMSE)
10.2 mm

Trajectory error (ATE RMSE)

We're looking for partners, collaborators, and early adopters.

Product overview

One platform, two deployment levels

NE-VSLAM ships in two editions that share a common software and ROS 2 interface. Start on the CPU you already have; move compute-intensive stages onto an FPGA when your latency or energy budget demands it.

NE-VSLAM Software Edition

Available now

Architecture-specific software optimization for embedded CPUs, with no discrete GPU or FPGA required.

Target platform

  • Raspberry Pi 4 Model B, 4–8 GB
  • Embedded ARM-based platform

Current result

  • Up to 31 FPS
  • Approximately 2.1× faster than the tested Stella-VSLAM and ORB-SLAM2 configurations
  • No discrete GPU or FPGA required

Optimization approach

  • Architecture-specific software optimization for embedded CPUs

NE-VSLAM FPGA Edition

In development

The same SLAM application, with selected compute-intensive stages offloaded to FPGA fabric.

Target platform

  • AMD Kria KV260
  • or other preferred FPGAs

Architecture

  • Embedded processor manages the SLAM application, robotics integration, and system control
  • FPGA accelerates selected compute-intensive processing stages
  • Common software and ROS 2 interface

Purpose

  • Higher throughput
  • More deterministic latency
  • Better efficiency for fixed workloads

Application scenario

Drops into the robot you already have

The same NE-VSLAM interface supports two deployment levels. The CPU-optimized version targets affordable robots and prototypes, while the CPU–FPGA version targets systems with stricter performance or energy requirements.

RGB-D camera

Input frames taken from an RGB-D camera.

Frames USB

Raspberry Pi 4 or Kria KV260

NE-VSLAM runs tracking, mapping and pose estimation on-device.

Pose (ROS 2) Ethernet

UAV or UGV

Your robot consumes pose directly for navigation and control.

Autonomous mobile robots (AMR)

On-board localization and mapping for indoor robot fleets.

Automated guided vehicles (AGV)

Reliable navigation for warehouse and factory transport.

Industrial & surveillance drones

Low-power, low-latency perception for battery-powered UAVs.

Research & prototype platforms

Affordable on-board SLAM for university labs and early robot builds.

Benchmark data

TUM RGB-D VSLAM benchmark comparison

NE-VSLAM delivers 31.255 FPS at 10.243 mm ATE RMSE, 2.104× the frame rate of ORB-SLAM2 and Stella-VSLAM.

ORB-SLAM2 and Stella-VSLAM are effectively tied at 14.860 and 14.858 FPS (0.010% mean-latency difference).

These are Software Edition results: CPU-only, with no FPGA or GPU acceleration involved.

Tracking latency

Y-axis: tracking latency (ms/frame) | lower is better | dashed line = 30 FPS target

0 15 30 45 60 75 ORB-SLAM2: 67.296 ms 67.296 ms (tie) ORB-SLAM2 Stella-VSLAM: 67.303 ms 67.303 ms (tie) Stella-VSLAM NE-VSLAM: 31.995 ms 31.995 ms NE-VSLAM Tracking latency (ms/frame)

Throughput

Y-axis: effective throughput (FPS) | higher is better | dashed line = 30 FPS target

0 7 14 21 28 35 ORB-SLAM2: 14.860 FPS 14.860 FPS (tie) ORB-SLAM2 Stella-VSLAM: 14.858 FPS 14.858 FPS (tie) Stella-VSLAM NE-VSLAM: 31.255 FPS 31.255 FPS NE-VSLAM Throughput (FPS)

Trajectory accuracy

X-axis: ATE RMSE (mm) | thin line = range; thick line = mean ± SD

9.40 9.60 9.80 10.00 10.20 10.40 ORB-SLAM2: mean 9.781 mm, SD 0.230, range 9.424–10.141 mm ORB-SLAM2 9.781 mm Stella-VSLAM: mean 10.249 mm, SD 0.196, range 9.912–10.502 mm Stella-VSLAM 10.249 mm NE-VSLAM: mean 10.243 mm, SD 0.144, range 9.955–10.407 mm NE-VSLAM 10.243 mm ATE RMSE (mm)

Speed–accuracy position

X-axis: effective throughput (FPS) | Y-axis: ATE RMSE (mm) | dashed line = 30 FPS target

9.6 9.8 10.0 10.2 10.4 15 20 25 30 30 FPS ORB-SLAM2: 14.860 FPS, 9.781 mm ORB-SLAM2 Stella-VSLAM: 14.858 FPS, 10.249 mm Stella-VSLAM NE-VSLAM: 31.255 FPS, 10.243 mm NE-VSLAM Effective throughput (FPS) ATE RMSE (mm)
ORB-SLAM2 Stella-VSLAM NE-VSLAM
TUM RGB-D Freiburg1 XYZ | Raspberry Pi 4 | 10 successful runs per system
View data as table
System Tracking latency Throughput ATE RMSE (mean ± SD) ATE range
ORB-SLAM2 67.296 ms 14.860 FPS 9.781 ± 0.230 mm 9.424–10.141 mm
Stella-VSLAM 67.303 ms 14.858 FPS 10.249 ± 0.196 mm 9.912–10.502 mm
NE-VSLAM 31.995 ms 31.255 FPS 10.243 ± 0.144 mm 9.955–10.407 mm

ROS 2 interface

Publishes pose straight into your ROS 2 graph

NE-VSLAM runs as a standard ROS 2 node. It subscribes to your existing RGB-D camera topics and broadcasts the map frame, so integration is a launch file rather than a porting project.

Subscriber TF static broadcast
/camera/rgb/image_color
sensor_msgs/Image
/camera/depth/image
sensor_msgs/Image

Camera

ne_vslam_rgbd
ROS 2 node

NE-VSLAM device

TF: map → odom
static transform

User robot

Custom acceleration service

Visual SLAM may not be your bottleneck

If your system has a computation, latency, or energy-efficiency problem, we can help optimize it, even when Visual SLAM is not the part that hurts.

Software

C++ optimization for your target CPU, including profiling, memory tuning, SIMD, and multithreading.

Hardware

Custom FPGA acceleration for AI, perception, and robotics workloads when CPU optimization is not enough.

Tell us about your robot, current computing platform, and main performance challenge. We will recommend the most suitable optimization path.

Describe your workload

Our story

Why we exist

NeuroEdge is a semiconductor research group founded in 2024 by alumni of ITS Surabaya. We develop hardware acceleration for real-time edge AI in robotics and autonomous systems. Autonomous robots need reliable localization when GPS and human control are unavailable or limited. Without proper optimization, CPUs can be too slow, while GPUs can use too much power. Our solution combines CPU architecture-specific software optimization with FPGA SoC acceleration for selected stages of the processing pipeline. We aim to close the gap between the perception algorithms developed by researchers and the hardware needed to run them efficiently on-device.

Our mission

Make advanced AI and 3D-vision algorithms run faster, cooler and cheaper on the devices where decisions actually happen.

A multidisciplinary team specializing in embedded software, FPGA design, digital integrated circuits, electronics, and robotics.

Lab B202, ITS Surabaya, Indonesia · Founded 2024

Contact

Tell us what you're building

We're looking for partners, collaborators, and early adopters. Share your robot, current computing platform, and main performance challenge. We will recommend the most suitable optimization path.

@ info@neuroedge.id +62 816 4818763
Lab B202, ITS Surabaya, Indonesia