falcondai / robot-grasp

a project to use learning to solve the robotic grasp problem

Robotic grasp

Use learning to solve the robotic grasp problem.

approach

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).

datasets

training the grasp evaluator

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

usage

use pre-trained models

training

authors

Falcon Dai ([email protected])