MuJoCo AR is a plugin for MuJoCo that enables the integration of ARKit data from a connected iOS device to control MuJoCo frames in real-time. It can also be used in non-MuJoCo applications (see the flexible setup section below). This package accompanies the MuJoCo AR iOS app that can be downloaded from here. Check the flexible setup if you wanna use it outside of MuJoCo. Finally, check the FAQ below.
- You can now trigger a vibration with configurable properties. This can be used to add haptic feedback, whether on contact with objects or to notify the user of specific events. More details in release log.
Position Control | |||
---|---|---|---|
MuJoCo Fruits Pick and Place | MuJoCo PushT | ||
Position and Rotation Control | |||
MuJoCo Study Desk | MuJoCo Blocks Stacking | ||
Examples of MuJoCo AR linked to the end-effectors of multiple manipulators can be found in this fork of Mink. For a simpler environemnt that the ones shwon in the demo section above, check the example by using mjpython demo/simple.py
-
iOS App: Download the iOS app from the App Store here
-
Python Package: Install the Python package using pip:
pip install mujoco_ar
This setup allows you to directly control a MuJoCo frame (body, geom, or site), with the frame's position and orientation matching the ARKit data received from the connected iOS device.
from mujoco_ar import MujocoARConnector
# Initialize the connector with your desired parameters
mj_ar = MujocoARConnector(mujoco_model=my_model,mujoco_data=my_data)
# Link a MuJoCo frame (link_body(), link_geom() or link_site())
mj_ar.link_body(name="eef_target")
# Start the connector
mj_ar.start()
In addition to what the quick setup allows you to do, this setup allows you to automate the applying of a translation, rotation or scaling of the recieved pose. Additionally, you can pass functions to button_fn and toggle_fn to be triggered when the button or toggle are activated
from mujoco_ar import MujocoARConnector
# Initialize the connector with your desired parameters
mj_ar = MujocoARConnector(
mujoco_model=my_model,
mujoco_data=my_data,
port=8888, # Optional, defaults to 8888 if not provided
debug=False # Optional, defaults to False if not provided
)
# Link a MuJoCo frame (link_body(), link_geom() or link_site())
mj_ar.link_body(
name="eef_target",
scale=1.0, # Optional, defaults to 1.0 if not provided
position_origin=np.array([0.0, 0.0, 0.0]), # Optional, defaults to [0, 0, 0] if not provided
rotation_origin=np.identity(3), # Optional, defaults to I(3) if not provided
toggle_fn=my_toggle_function, # Optional, calls nothing if not provided
button_fn=my_button_function, # Optional, calls nothing if not provided
disable_pos=False, # Optional, defaults to False if not provided
disable_rot=False # Optional, defaults to False if not provided
)
# Start the connector
mj_ar.start()
You can retrieve the ARKit data including the position, rotation, button, and toggle states directly from a connected iOS device, making it flexible for usage in various applications beyond physics simulations. Try running mjpython demos/flexible_setup.py
.
from mujoco_ar import MujocoARConnector
# Initialize the connector with your desired parameters
connector = MujocoARConnector()
# Start the connector
connector.start()
# Retrieve the latest AR data (after connecting the iOS device, see the guide below)
data = connector.get_latest_data() # Returns {"position": (3, 1), "rotation": (3, 3), "button": bool, "toggle": bool}
connector.vibrate(sharpness=0.8, intensity=0.4, duration=0.01) # Trigger a vibration on the connected device
connector.pause_updates() # Temporarily stops receiving updates from the connected device.
connector.resume_updates() # Resumes receiving updates from the connected device.
connector.reset_position() # Resets the current position as the origin (0,0,0).
- If you're experiencing latency, try connecting your PC to your device's hotspot. This should significantly reduce latency if you're far from a router since the communication happens locally via WebSockets.
- Make sure you're not covering the camera with your fingers. If the tracking is too sensitive, you can adjust the
scale
attribute when linking the frame to make it less reactive to minor movements. In my experience, the position retrived is more accurate than the rotation.
- I highly recommend running
demo/simple.py
to test things out if you're just getting started. This demo is configured to change the sphere's color with a button press and enlarge it with a toggle, providing a basic but effective introduction to a simple configuration.
- Yes, check the Flexible Setup out where you can retrive the pure ARKit position and rotation and use it as you wish. You wouldn't need to pass in the MuJoCo model and data in such a case.
Report any bugs you encounter to the issue tracker. Also, feel free to suggest any features. Those could be regarding the app (iOS swift code) or the python package.
Thanks to @kevinzakka for the mink and mjctrl libraries, which are used in the provided demos.
Omar Rayyan (olr7742@nyu.edu)