Tracking Issue on ALOHA Performance improvement using dora-rs

[haixuanTao]

Performance

• Rust rewrite of teleop for lower than 20ms latency

  • one_side_teleop.py seems to not be super efficient, and has a DT=20ms. It seems that 20ms is noticeable so maybe try to aim for like 10ms?
  • Rust can deliver on multithreaded-concurrent subscribing and publishing

• Faster Webcam latency, enable real-time visualization? Faster End-to-End Inference?

  • sudo apt-get install ros-noetic-usb-cam && sudo apt-get install ros-noetic-cv-bridge This depency should be less efficient than using cv2 from dora-rs IMO but need to try it out. Also could be hard to install on certain computers.

Tooling

• Move matplotlib to rerun.io

  • visualize_episodes.py this script can probably also be optimized using a dora-rerun node, using also multithreaded concurrent visualization.

• Multithreaded concurrent recording

  • record_episodes.py this script should be optimisable using dora-record, albeit tweak dora-record to record ROS/ROS2 messages

Ergonomy

• Enable more than Python 3.8 and Ubuntu 20.04
  • conda create -n aloha python=3.8.10 Also this is is not major, I feel like this could be a very big frustrating point
  • Ubuntu 20.04 This could be a big pain point as not everyone can switch OS easily. So fixing this can enable mass-audience.

Hardware

• Use Gripper with GoPRO as proxy for teleoperation of smaller arms.

[Hennzau]

Regarding your "Ergonomy points," it would be great to create a Docker image with everything needed to run dora-lerobot on Aloha Arms:

• Ensure compatibility with multiple operating systems.
• Enable the creation of images for multiple architectures.
• Ensure that the Python environment is correctly installed.

I can try to make it, tell me if it's good or if it's too early

[haixuanTao]

I mean, i think dora-lerobot should be natively cross-platform and multi-architecture and it already is I think.

A good conda, pip or poetry environment description could really help but I don't think that we should spend to much time on docker images as well as maybe a quick cargo install description.

I know that docker image can be easily failure prone and hard to maintain.

Also, peolpe might already use docker and we don't to do docker in docker

[Hennzau]

I see! I'll probably do something about env description