A System for General In-Hand Object Re-Orientation

In-hand object reorientation, i.e., choosing an object, reorienting it in hand, and then positioning it or utilizing it as a software, is a popular maneuver for humans. Even so, this task in robotics remains an lively spot of research because of to several unsolved challenges.

Object re-orientation is a difficult task for machines.

Item re-orientation is a challenging task for equipment. Picture credit: Jeremy Tarling via Wikimedia, CC-BY-SA-two.

A latest paper on arXiv.org proposes a technique for the object reorientation issue with a multi-finger hand. It permits manipulation with arms dealing with upward or downward and is able to use external surfaces to assist manipulation. What’s more, the technique can reorient objects of novel designs to arbitrary orientations. It operates from sensory info that can be quickly obtained in the true planet.

In get to attain that, the technique works by using product-absolutely free reinforcement finding out with 3 important parts: teacher-university student finding out, gravity curriculum, and steady initialization of objects. The researchers provide evidence that the uncovered procedures can be transferred to the true planet in the long run.

In-hand object reorientation has been a complicated issue in robotics because of to large dimensional actuation area and the repeated modify in speak to point out concerning the fingers and the objects. We present a uncomplicated product-absolutely free framework that can find out to reorient objects with both equally the hand dealing with upwards and downwards. We reveal the ability of reorienting above 2000 geometrically distinct objects in both equally conditions. The uncovered procedures clearly show potent zero-shot transfer general performance on new objects. We provide evidence that these procedures are amenable to true-planet procedure by distilling them to use observations quickly accessible in the true planet. The video clips of the uncovered procedures are accessible at: this https URL.

Study paper: Chen, T., Xu, J., and Agrawal, P., “A System for Normal In-Hand Item Re-Orientation”, 2021. Website link: https://arxiv.org/abs/2111.03043