This remarkable paper shows that by increasing data (i.e., grasp attempts) by 40x compared to the largest prior work, one can continually run a robot to attempt grasps and eventually train it to grasp objects reliably, including novel objects not seen during testing.
Key technical contributions:
-
Self-supervision, where objects (in CLUTTER) are placed on the workspace, and the (Baxter) robot continually tries to pick up objects.
-
A simple "curriculum" or "staged learning" phase when, after some random exploration/actions, the robot uses a trained neural network and then samples actions based on that. Hard negatives (i.e., inaccurate grasps predictions, sort of) are used to assist in training by providing greater signal.
-
The neural network takes in a camera image, processes it through a set of pre-trained AlexNet layers, and then performs an 18-way binary classification problem, because there are 18 different parallel-jaw grasp angles (discretized into bins) and multiple angles could lead to successful grasps. You can apply 18 softmaxes, one per discretization, if you want a probabilistic interpretation. Also, classification is easier than regression.
Update 09/2018: wow, how did I not see this before? It's not the camera image but a patch that is within the larger camera sensor image. The patches are centered about an object --- which one can find by simple Mixture of Gaussian or other non-learning methods --- and then we get a bunch of angles about the patch center.
-
Grasping is done entirely by vision-based judgment, not an analytic model of 3D objects, etc. No human labels as well, either. The labels come from measuring the forces on the gripper after gripping, which is similar to what Sergey did, and works well for objects that are not too thin (which would be hard to grasp anyway).
This paper shows the remarkable ability to learn grasp policies that actually generalize to novel objects outside the training data.
How does this relate to Sergey's arm farm paper? It was the predecessor to it. It doesn't have the continual servoing mechanism that Sergey's paper has and requires camera calibration, but this one has a lot of what the later one does, i.e. just learning from tries with automatic data labels. Also, the arm farm paper did not use the "discretize angles into bins" approach as this one does.
Potential limitations? It's data hungry, I suppose. But this work is very impressive, so hard for me to criticize ... unless one just doesn't like self-supervision papers (the critics, heh)? In terms of writing, I think the experimental section needs to be improved to match the quality of the rest of the paper, it's lacking some clarity. For example can the 79.5% figure be derived by the stuff in Figure 8, which shows other accuracy results? It doesn't look like it.
IMPORTANT NOTE ON EVALUATION: for testing, they can execute grasps with the robot by sampling a set of patches and determining the best patch+angle combination. However, their network is evaluated based on whether it correctly predicted the outcome (success or failure). I got confused by this originally, because I thought this was purely about success rates in practice (i.e., it grasped 80% of objects successfully), but they actually measure accuracy. That's somewhat disappointing, actually, as I wanted to see actual success. They do report that in Section IV-F though it seems to be less heavily emphasized as the other parts of the paper. Rates are roughly 66% to 73%, not bad!