Efficient Visual Avoidance and Control for UAVs with Embedded GPUs
summary
A compact quadrotor runs a full visual navigation stack onboard a Jetson Tegra X1 (companion computer) and Pixhawk flight controller, doing stereo depth via GPU-accelerated semi-global matching, visual-inertial odometry at 30 Hz and reactive avoidance with a 30 ms capture-to-avoidance latency. The system flies 20 minutes, keeps centimeter-level depth precision at ~1–4 m, uses hardware-synced global-shutter cameras and CUDA kernels to hit real-time rates, and maintains a 5 m local map for fast reactive avoidance and GPS-denied indoor hover.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:39STATE OF THE INDUSTRY
- 5:18PROJECT ARTEMIS - A HISTORY
- 10:54DESIGN GOALS
- 12:11VEHICLE PLATFORM - Artemis MAVS
- 13:06VEHICLE SPECS
- 13:48TEGRA X1 COMPUTE PLATFORM
- 14:47THE NAVIGATION PROBLEM
- 16:42SENSING SUITE Framerate
- 20:20VISUAL-INERTIAL ODOMETRY
- 23:42VISUAL-INERTIAL STATE ESTIMATION
- 34:46GLOBAL PLANNING
- 37:05NONLINEAR OPTIMAL CONTROL
- 38:24OPERATOR INTERFACE
- 39:19SOFTWARE FRAMEWORK
- 40:30COMPLETE NAVIGATION PIPELINE
- 42:38MULTI-SENSOR FUSION
related talks
13mBlind Spot in Sensor Fusion: Exploiting an Architectural Vulnerability to Hijack...
21mBitDrones: Towards Using 3D Nanocopter Displays as Interactive Self-Levitating Programmable Matter
30mAI On-board of Satellites, Towards Autonomous Scientific Instruments
06mTiny Flying Robots
46mPaparazzi, An Open Source UAV for Research
47mNext-gen Dynamic UAV using Power9 Systems
08mAR for Flight Control at AWE 2015
27mAutomated Python Powered Turret