VisionSense™
An open-source perception platform for autonomous driving. Runs seven deep learning models concurrently on NVIDIA Jetson, powered by integrated stereo vision, IMU, and GNSS.
Drag to rotate · Scroll to zoom
Platform Architecture
Jetson Orin Nano / NX
67 / 117 / 157 TOPS edge AI performance with 8GB / 16GB LPDDR5 memory tiers running concurrent perception models.
Multi-Camera Array
Hardware-synchronized cameras pre-calibrated for reliable stereo depth estimation.
Navigation & IMU
Integrated spatial orientation, absolute positioning, and vehicle odometry tracking.
Snap-On Mount Enclosure
3D-printable protective enclosure featuring quick-release magnetic vehicle mounts.
4G/LTE Module
Adds a plug-and-play global SIM carrier for remote diagnostics and cloud OTA updates.
Open-source ROS2 Software
Zero-setup Docker container pre-loaded with inference pipelines, visualization nodes, and active camera drivers.
Built for Autonomy Teams
University Labs
Equip your research group with an affordable, ready-to-run platform to reproduce publications and record multi-camera datasets.
AV Startups
Accelerate road testing and validate camera-compute configurations instantly before committing to production hardware.
Fleet Operators
Deploy road condition surveys and driver telemetry using custom mounts and remote LTE connectivity.
Robotics & Makers
Build scaled autonomous test beds, facility inspection rovers, and customized sensor setups.
ROS2 Perception Dashboard
Interactive recreation of the VisionSense dashboard. Real-time telemetry, multi-stream switching, and live detection logs. This is what runs on the device.
Real-time inference on device
The full stack runs concurrently. No cloud dependency, no network latency. Every model is TensorRT-optimized for the selected Jetson module.
Numbers measured on Orin Nano Super at 1080p. Performance varies by scene complexity, lighting, and compute tier selected.
| Task | Performance |
|---|---|
| Object Detection | 54-71% mAP@50 |
| Multi-Object Tracking | Real-time |
| Lane Detection | 81.5% F1 CULane |
| Traffic Sign & Light Classifier | 96% accuracy |
| Stereo Depth Estimation | 0.3m - 50m range |
| Face Detection | 92.5% mAP@50 |
| Gaze Estimation | 4.5° angular error |
Hardware and software,
all in the open.
The repository contains everything: 3D-printable enclosure STLs, the full ROS2 perception pipeline, all trained model weights, and TensorRT optimization scripts. Fork it, build on it, contribute back.
- Enclosure STL files and STEP source
- Pre-trained TensorRT models and training scripts
- ROS2 perception pipeline (Humble / Iron / Jazzy)
- Regular model releases and community contributions
# VisionSense Perception Models
# All models TensorRT-optimized for Jetson Orin
models:
object_detection:
architecture: YOLOv13
weights: weights/yolov13_medium.engine
precision: FP16
classes: 10
lane_detection:
architecture: CLRerNet
weights: weights/clrernet_culane.engine
precision: FP16
traffic_signs:
architecture: MobileNetV4
weights: weights/mobilenetv4_signs55.engine
classes: 55
precision: INT8
stereo_depth:
architecture: LightStereo
weights: weights/lightstereo.engine
max_disparity: 192
precision: FP16
tracking:
algorithm: BoT-SORT
max_age: 30
traffic_lights:
architecture: YOLOv13
states: [red, yellow, green, arrow]
precision: FP16
driver_monitoring:
architecture: YOLOv11 + ResNet18
precision: FP16What's included in your system
Everything you need to begin deployment. Plug in, power up, and access the ROS2 nodes immediately.
VisionSense Head Unit
Configured Jetson module, cameras, IMU, and GNSS pre-assembled inside a carbon-fiber ASA enclosure with free-use design license.
Open-Source ROS2 Pipeline
Pre-installed Humble/Jazzy perception nodes, trained model weights, and the live AutoVision web dashboard.
Snap-On Magnetic Mount
High-strength N52 neodymium magnetic snap-on base with an adhesive steel mount plate for simple roof/windshield installation.
External GPS Antenna
Active, rugged external car-mount GPS antenna with SMA connection for precise sub-meter localization.
Power & Video Cables
Includes high-speed USB-C power delivery cable, DisplayPort to HDMI cable, and an outdoor-grade Gigabit Ethernet cable.
Developer SDK & Guide
Access to comprehensive configuration manuals, source CAD/STEP mechanical designs, and hardware calibration files.
Join the open perception community
Researchers and engineers worldwide are building with VisionSense. Open hardware, open software, active development. Our team is reachable on GitHub.