Learning Object Manipulation Skills via Approximate State Estimation from Real Videos

Vladimír Petrík*, Makarand Tapaswi*, Ivan Laptev, Josef Šivic

CIIRC, Czech Technical University in Prague Inria


This page presents supplementary materials for our CoRL 2020 submission on Learning Object Manipulation Skills via Approximate State Estimation from Real Videos. In case of any question contact us at vladimir.petrik@cvut.cz or makarand.tapaswi@inria.fr.

Paper arXiv Video Source code

Real2Sim results

We visualize results as GIFs showing all 6 videos for each of the 9 actions. Each row shows a full video, followed by segmentation masks and corresponding 3D state renderings via the neural renderer.

View details »

Policy training results

For each action, we present one successful and one failure example in the simulator, and the result of transferring the policy to a real robot.

View details »

Intermediate results

Intermediate results allow you to run individual parts of the proposed pipeline. For example, if you are primarily interested in RL, you can start with our estimated states without running trajectory optimization.

Download segmentation masks

Download estimated states