Robots & Sensor Suites

Due to the nature of the environments where we captured the EnvoDat dataset, we used two robot platforms. For indoor environments with structured layouts, we utilised our open-source wheeled robot (ROMR).  However, for off-road and outdoor environments with unstructured terrains, we utilised our Unitree Go1 quadruped robot. Our sensor setup on the robots is illustrated in the figure below. 

We calibrated with a 300 x 390 mm chessboard with 15 x 10 internal corners of 21 mm square length. During calibration, 20 image samples were captured at approximate left-right rotation angles of 0°, 45°, 90°, 135°, and 180°, and front-back bending angles of 30°, from distances between 1 to 3 meters. We intentionally did not exceed this range since the monocular cameras for example have a very low resolution, which when exceeded may lead to inaccurate detection of the chessboard corners. We obtained the intrinsic and extrinsic parameters of the sensors using the following tools:

Cameras:

• We used the Kalibr (https://github.com/ethz-asl/kalibr)  target-based calibration pipeline and the 

• MATLAB calibration App (https://uk.mathworks.com/help/vision/camera-calibration.html). 

Camera-LiDAR:

• MATLAB calibration App (https://uk.mathworks.com/help/vision/camera-calibration.html). 

• Kalibr (https://github.com/ethz-asl/kalibr).

IMU:

• Allan Variance (https://github.com/ori-drs/allan_variance_ros) is used to compute the angle, velocity random walks as well as the bias instabilities of the accelerometer and the gyroscope. 

Camera-IMU:

• We utilised the Kalibr (https://github.com/ethz-asl/kalibr) visual-inertial calibration toolbox to perform spatial and temporal calibration of the IMU with our ELP 4K Megapixel monocular camera.