This content was created by LARRUZ. This content was uploaded by LARRUZ. This is a YouTube video titled "This Robot Learned To Clean Up Clutter" (created 2018-10-07) consisting of a monologue by the Two Minute Papers host. The narrator describes a robotics research demo in which a robot arm equipped with an RGB-D camera learns to clear clutter by combining pushing and grasping actions. The system uses two neural networks (one predicting pushing utility, one for grasping) and chooses motions that most improve the table state. The video explains training in simulation to accelerate learning, comparisons to state-of-the-art, and the method's generalization to novel objects—framed as a step toward agents that perceive and plan in the real world.
Chapters
Dear fellow scholars, this is two minute papers. With this robot was tasked to clean up this table. Normally, anyone who watches this series knows that would be no big deal for any modern learning algorithm. Just grab it, right? Well, not in this case, because reason number one, several objects are tightly packed together. And reason number two, they are too wide to hold with the fingers. What this means is that the robot needs to figure out a series of additional actions to push the other pieces around and finally grab the correct one. Look, it found out that sometimes pushing helps grasping by making space for the fingers to grab these objects. This is a bit like the Roomba vacuum cleaner robot, but even better for clutter. Really cool. This robot arm works the following way. It has an RGBD camera which endows it with the ability to see both color and depth. Now that we have this image, we have not one, but two neural networks looking at it. One is used to predict the utility of pushing at different possible locations, and one for grasping. Finally, a decision is made as to which motion would lead to the biggest improvement in the state of the table. So what about the training process? As you see, the speed of this robot arm is limited and we may have to wait for a long time for it to learn anything useful and not just flail around destroying other nearby objects. The solution includes my favorite part, training the robot within a simulated environment where these commands can be executed within milliseconds, speeding up the training process significantly. Our hope is always that the principles learned within the simulation occ applies to reality. Checkmark. The simulation is also very useful to make comparisons with other state of the art algorithms easier. And do you know what the bane of many, many learning algorithms is? Generalization. This means that if the technique was designed well, it can be trained on matte looking wooden blocks and it will do well when it encounters new objects that are vastly different in shape and appearance. And as you see on the right, remarkably, this is exactly the case. Checkmark. This takes us one step closer to learning algorithms that can see the world around us, interpret it, and make proper decisions to carry out a plan. Thanks for watching and for your generous support and I'll see you next time.
