Visual SLAM for aerial manipulation

Problem

LAAS-CNRS estimates its aerial robots’ state by fusing the on-board IMU with a motion capture system in an Unscented Kalman Filter. That works indoors, but motion capture is unusable outdoors and its bandwidth makes fully wireless operation impractical. The goal was a vision-based alternative good enough for a quadrotor that carries an arm and physically interacts with the world, where every gram of payload and every watt counts.

What I did

During a two-month internship in Dr Antonio Franchi’s Robotics and InteractionS team, I worked on vision-based state estimation. We chose the Intel Euclid, a 146 g self-contained active-stereo camera with its own computer and battery, so the vision system needed no extra wiring or share of the motor power. I integrated and tuned ORB-SLAM2 to run on the Euclid alongside its built-in RealSense SLAM, briefly evaluated ROVIO and OKVIS as alternatives, brought the data into MATLAB/Simulink, and designed a 3D-printed test bed for the sensor. I then compared both SLAM systems against motion capture ground truth over three trials.

My second task was the gripper. I surveyed commercial grippers and selected the Makeblock gripper for its 67 mm opening, light weight and self-locking fingers, then soldered the electronics (Arduino Nano, motor driver, 9 V DC-DC converter) and wrote the Simulink control that opens and closes it to match the width of the object. I also assembled two quadrotors, one with the 2-DoF arm and gripper.

Result

On board the Euclid, ORB-SLAM2 tracked position significantly better than RealSense SLAM, while RealSense SLAM was generally better on orientation. ORB-SLAM2 occasionally lost track and relocalised, likely because it is heavy for the Euclid’s small computer, and fast, abrupt motion in the third trial degraded both. The gripper control, calibrated on 4 cm and 3 cm bars, gripped and released an untested 3.5 cm bar as predicted.