This code was partly used for a robot competition.

Contains scripts for sim2real segmentation and RL for top-down grasping.

Todo: better documentation, including data preparation (which includes ShapeNet models)

Use https://github.com/sea-bass/ycb-tools to download YCB models.

Place models/ycb in repository (i.e. hsr_pybullet/ycb/002_master_chef_can...)

train_agent.py is an example of RL for grasping

tmc_wrs_gazebo, which contains assets for the room based on that of the competition, was copied from https://github.com/hsr-project/tmc_wrs_gazebo/tree/master/tmc_wrs_gazebo_worlds.

hsrb_description, which contains the URDF of the Toyota HSR robot, was copied from https://github.com/hsr-project/hsrb_description.

hsrb_meshes, which contains the meshes referenced by the robot URDF, was copied from https://github.com/hsr-project/hsrb_meshes.

Part of the panoptic segmentation code comes from https://github.com/bowenc0221/panoptic-deeplab.

The depth sensor noise model comes from https://github.com/facebookresearch/habitat-sim.

A subset of 3D models from the ShapeNet dataset are used by this project. Due to restrictions, these models are only downloadable from the official ShapeNet dataset website.