Q-attention: Enabling Efficient Learning for
Vision-based Robotic Manipulation

Stephen James Andrew J. Davison

Dyson Robotics Lab, Imperial College London

Code: https://github.com/stepjam/ARM

Trained Policies (with Q-attention visualised)

Videos show the Q-attention "heatmap" along with the chosen crop.

qattention_put_rubbish_in_bin.mp4

put_rubbish_in_bin

qattention_take_lid_off_saucepan.mp4

take_lid_off_saucepan

Keyframe Discovery

Visualisation of the keyframe discovery for 4 tasks.

put_rubbish_in_bin_keyframes.mp4

put_rubbish_in_bin

reach_target_keyframes.mp4

reach_target

stack_wine_keyframes.mp4

stack_wine

take_lid_off_saucepan_keyframes.mp4

take_lid_off_saucepan

Example Demonstrations

example_demos.avi