Reproduce “Learning to Place Objects onto Flat Surfaces in Upright Orientations” on OpenResearch

About this paper

Rhys Newbury, Kerry He, Akansel Cosgun, Tom Drummond

This research from Monash University introduces an iterative deep learning method that enables a robot to place grasped objects onto flat surfaces in their semantically preferred upright orientations. This approach achieves 98.1% success in simulation with unseen objects and 88.0% on a physical robot, using only depth images without prior object models or class information.

Key insights

• Iterative refinement, combined with a quality assessment network, substantially improves placement success, effectively correcting common failure modes like placing objects upside-down.
• A 6D continuous representation for rotations and pre-training the CNN backbone on ImageNet are critical for achieving high performance and accurate rotation prediction.
• Deep learning can successfully capture human semantic preferences for object "uprightness" and generalize these to unseen objects using only visual input, bypassing the need for explicit 3D models or object class labels.

Summary by alphaXiv.