The project RobotFlow aims to design a new kind of neuro-robotics system based on organic controllers and bio-inspired vision robotics for better performance, automation and the minimization of sensors.

Dataset with RGB-images and sensor-based distances to objects

This dataset comprises of 3,431 tuples of simulation data acquired using V-REP. Each tuple contains:

• images - 256 x 256 RGB images acquired from an egocentric view monocular camera
• images_gray - 256 x 256 grayscale images converted from images
• distances - closest distance from robot-to-object measured with an ultrasonic sensor, in meters
• velocity - V-REP value for joint controllers representing the speed of motor rotation
• time - simulation time of each sample, in milliseconds

[download] ConvNet_Data.mat - MATLAB file with the dataset

Dataset with optical flow calculations and sensor-based distances to objects

This dataset comprises of 860 tuples of simulation data acquired using V-REP and optical flow calculations. Each tuple contains:

• images_uv - 256 x 512 grayscale images corresponding to the concatenation of the horizontal and vertical components of optical flow acquired between two images (at times t and t + 4 respectively) from images in the previous dataset
• distances - closest distance from robot-to-object measured with an ultrasonic sensor, in meters
• velocity - V-REP value for joint controllers representing the speed of motor rotation
• time - simulation time of each sample, in milliseconds

[download] ConvNet_Data_Flow_Augmented.mat - MATLAB file with the dataset

Videos of Experimental Results

Here you will find several videos recorded from simulations using the neuro-robotics system based on CNN and optical flow reported in the paper.

1. attempt 1 using CNN+OF - Video file
2. attempt 2 using CNN+OF - Video file
3. attempt 3 using CNN+OF - Video file
4. attempt 4 using CNN+OF - Video file
5. attempt 5 using CNN+OF - Video file
6. attempt 1 using CNN without OF - Video file