UC Berkeley
Code: https://github.com/stepjam/ARM
(Repo also contains code for ARM and C2F-ARM)
C2F-ARM has already been qualitatively evaluated in the real world, and has demonstrated the ability to be trained efficiently from scratch.
To highlight the usefulness of the tree expansion in the real world, we design a real-world task that involves three small yellow objects with different shapes, such that only at the finest levels of the Q-attention would they be able to be distinguishable.
The task is to simply reach the Pentagonal prism, while ignoring the two distractors. We give both C2F-ARM and our C2F-ARM+QTE three demonstrations (through tele-op via HTC Vive), and train for 10 minutes. When evaluated on 6 test episodes, C2F-ARM was only able to reach the target 2/6 times, while C2F-ARM+QTE achieved 6/6.
The 4 tasks from 'Task Set 2'. These were chosen due to their ambiguity at a coarse scale.