Use learning to solve the robotic grasp problem.
We factored the system into 3 components:
- proposer: take in an image (RGB+D) then propose a grasp to be evaluated
- evaluator: score the grasps
- executor: execute the recommended grasps
and we will learn/optimize each component using data collected on and off a robot (Baxter).
To train a reasonable grasp evaluator, we treated the evaluation problem as a binary classification task and used the publicly available Cornell grasp dataset to train off the robot. Here are some of statistics of this dataset:
- image count: 885
- labeled grasps count: 8019
- positive: 5110 (64%)
- negative: 2909 (36%)
- object count: 244
- object category count: 93
Falcon Dai (dai@ttic.edu)