Dual SERL

For detailed information about this project, please refer to the paper included in this repository:
Oliani_two_arms_one_goal.pdf.

Evaluations

Contributions

Code Directory	Description
robot_controllers	Impedance controller for the UR5 robot arm
dual_ur5_env	Environment setup for the UR5 env
vision	Point-Cloud based encoders
utils	Point-Cloud fusion and voxelization
her	Hindsight Experience Replay for online learning

Quick start guide for box picking with a UR5 robot arm

Without cameras (TODO modify the bash files)

1. Follow the installation in the official SERL repo.
2. Check envs and either use the provided box_picking_env or set up a new environment using the one mentioned as a template. (New environments have to be registered here)
3. Use the config file to configure all the robot-arm specific parameters, as well as gripper and camera infos.
4. Go to the box picking folder and modify the bash files run_learner.py and run_actor.py. If no images are used, set camera_mode to none . WandB logging can be deactivated if debug is set to True.
5. Record 20 demostrations using record_demo.py in the same folder. Double check that the camera_mode and all environment-wrappers are identical to drq_policy.py.
6. Execute run_learner.py and run_actor.py simultaneously to start the RL training.
7. To evaluate on a policy, modify and execute run_evaluation.py with the specified checkpoint path and step.