Researchers led by Rotem Atari have shown that a robot can estimate the full 3D shape of a deformable linear object from touch and its own joint positions, without any camera, according to a paper on arXiv. Vision often fails with such flexible objects because they get hidden and can take countless shapes.

Their physics-informed vision transformer turns dense, high-resolution tactile images into a chain of 3D points, and a loss weighted by physics keeps the predicted shape physically plausible.

The touch-based method reconstructed shapes more accurately by both Euclidean and Chamfer measures, and it transferred to new materials and shapes of object. The abstract gives no figures.

Source: arXiv Robotics (cs.RO)

Background

Rotem Atari

Robotics researcher

Robotics researcher.