Autonomous robotic manipulation comes from an effective use of materials, modeling, perception, planning, control and design. The MCube Lab brings those disciplines together for the problem of mastering contact. Our goal is to develop technologies to bring reliable autonomous physical interaction closer to reality.
Information about accessibility can be found at accessibility.mit.edu/
Impressive work, I enjoyed your video. I wish you guys would have included failure cases in your video. It would be informative to know the limits of your robotic arm.
I'm always amazed at how complex even picking up and moving an object is. I like the combination of neural net (CNN) and pose estimation of the pre-scanned object. Wonder how far we are from having an end-to-end neural net, RGB to joint torques to achieve the goal. Of course you'd give up any kind of control over grasps/etc so this current solution is nice. Great work all!
Cool! Perhaps a model that can reevaluate its predictions every second or so given new data of the object being pushed would suffer less from poor long term predictions, and the short term predictions would be adequate.
