Intro:

This package provide a simple Franka arm and Robotiq Gripper simulator written in Mujoco. It includes a state-based and a vision-based Franka lift cube task environment.

Installation:

• From hil-serl folder, cd into franka_sim.
• In your hil-serl conda environment, run pip install -e . to install this package.
• run pip install -r requirements.txt to install sim dependencies.

Explore the Environments

• Run python franka_sim/test/test_gym_env_human.py to launch a display window and visualize the task.
• Run python franka_sim/test/test_gym_env_joystick.py --controller_type=<controller_type> to launch a display window and use joystick to control the arm. controller_type can be ps5 or xbox. You can use the keyboard to switch the camera view of each window by pressing [ or ].

Run Experiments

• Run python examples/record_success_fail_sim.py --exp_name pick_cube_sim --successes_needed 2000 to record success and failure trajectories which is used for training the reward model, you can define the number of successes needed. Controller type can be set in examples/experiments/pick_cube_sim/config.py. You can use the keyboard to switch the camera view of each window by pressing [ or ].
• Run python examples/record_demos_sim.py --exp_name pick_cube_sim --successes_needed 30 to record demonstrations for training the policy, you can define the number of successes needed. Controller type can be set in examples/experiments/pick_cube_sim/config.py. You can use the keyboard to switch the camera view of each window by pressing [ or ].
• To train a rlpd agent to solve the pick cube task:
  • cd examples/experiments/pick_cube_sim
  • run bash run_actor.sh, Update the command-line arguments in run_actor.sh based on your specific settings.
  • run bash run_learner.sh, Update the command-line arguments in run_learner.sh based on your specific settings.

Troubleshooting

Sometimes the window will frozen and simulation will stop. Even if you killed the process, the program will still be running in the background. You can kill the process by running

• kill -9 $(pidof simulate).
• lsof /dev/nvidia* to find the process id of the program using the GPU and kill the relevant process.

Credits:

• This simulation is initially built by Kevin Zakka.
• Under Kevin's permission, we adopted a Gymnasium environment based on it.

Notes:

• Error due to egl when running on a CPU machine:

export MUJOCO_GL=egl
conda install -c conda-forge libstdcxx-ng