mathildekappel/trajectory_primitive

Overall presentation of QDTraj Object-Centric Trajectory Dataset

This repository contains a dataset of object-centric cartesian trajectories generated with QDtraj. Each trajectory represents the motion of the FR3 end-effector expressed in the object coordinate frame. If you want to quickliy visualize outcome archive exemples, please go the following hugging face space : trajectory_primitive_visualisation_space. See QDtraj paper and QDtraj website for more details about the evolutionary algoritm used for the generation.

Structure

The Hugging Face repository contains two folders:

• comparison_study: This folder contains the results of the comparison study lead in the paper.
• trajectory_database: This folder contains trajectories generated using the best-performing combination (Adaptive interaction policy and MAP-Elites Explore) for many different objects from the PartNet-Mobility dataset, as well as for our experimental objects.

Independant outcome archives

This dataset provides independ folders of expert trajectories. Each outcome archive folder stores diverse trajectory allowing to activate a specific activation primitive. Here is the following folder convention for outcome archive storage: IdObject_TypeJoint_IndexJoint (ie. independent_folder )

Content of one outcome folder

All experts trajectories of one outcome archive are stored in a Archive_space_Object_..._JointNature_..._Item..._Generation_..._...PopSize_..._baseline_..._initState... .jsonl ie. Archive_space_Object.. .jsonl.

The meta-data associated to an outcome archive is stored in ... _time.jsonl file. (ie. Archive_space_Object.. ._time.json.)

You can easily visualize best trajectories of the outcome archive folder using the .mp4 (or .gif) videos located inside each folder. (ie. video_results_archive... .mp4 )

Visualisation an replay

You can easily replay trajectories of the outcome archive folder using the replay...mp4 (or .gif) videos located inside each folder. Trajectory archive folders can be visualized and replayed in Genesis simulation (see QDTraj source code readme for more details). You can also visualise and deploy the expert trajectories on your real FR3 robot (see Integration QDTraj source code readme)

Genesis Replay (QD-Trajectory Viewer)

Sotre the outcome archive in other_pkgs/resutls/ and run the replay mode.

contact

kappel@isir.upmc.fr