Add camera-controlled Panda demo with YOLOv8 person detection

[ChadliaJerad]

This PR:

• Adds PandaDemoCamCtrl.lf: Panda robot that pauses automatically when ≥2 people are detected by a YOLOv8 webcam feed, and resumes when the area is clear
• Migrates video library from YOLOv5 to YOLOv8 (faster on CPU, simpler API)
• Adds a dataset to PandaDemoParameters.csv with joint limits tuned to keep the arm above the floor
• Updates source README and adds src/lib/video/* files with setup and run instructions

[hokeun]

@ChadliaJerad @edwardalee I was explaining this demo to my student today and noticed that this was not merged yet. Could we merge this PR into main soon?

[ChadliaJerad]

Sure! How about Edward doing the review, me working on the comments he will give, and you proceeding with the merge?

[hokeun]

Sure! How about Edward doing the review, me working on the comments he will give, and you proceeding with the merge?

Sounds good to me, thanks!

[edwardalee]

I am unable to get this working. See comments. I'm not crazy about the duplication of the YOLO code with the playground repo. It would be better to find a package manager solution so that both playground and this could import the video library. Or maybe this should go into the playground, which should import this mujoco-py library.

[edwardalee]

Why not auto-detect whether there is a NVDIA GPU, Apple GPU, or CPU? Here is a possible template:

pip install torch torchvision

Then the following code will detect the device and use it in a test:

import torch

def get_compute_device() -> torch.device:
    """
    Detects and returns the best available computing device.
    Prioritizes CUDA, then Apple MPS, and falls back to CPU.
    """
    if torch.cuda.is_available():
        device_name = "cuda"
    elif torch.backends.mps.is_available():
        device_name = "mps"
    else:
        device_name = "cpu"
        
    print(f"🚀 Device selected: {device_name.upper()}")
    return torch.device(device_name)

if __name__ == "__main__":
    # 1. Initialize the device
    device = get_compute_device()
    
    # 2. Test tensor creation directly on the active device
    try:
        x = torch.randn(3, 3, device=device)
        y = torch.randn(3, 3, device=device)
        z = torch.matmul(x, y)
        
        print("\n✅ Verification Test Passed!")
        print(f"Matrix multiplication result shape: {z.shape}")
        print(f"Tensor is successfully hosted on: {z.device}")
        
    except Exception as e:
        print(f"\n❌ Error during tensor operations: {e}")

[edwardalee]

If we auto-detect, then requirements.txt should include these.

[edwardalee]

On a mac, I had to do this:

mjpython src-gen/PandaDemoCamCtrl/PandaDemoCamCtrl.py

Then, the MuJoCo window opens, but then I get this error:

ERROR: FATAL: Calling reaction _display.reaction_function_1 failed.
Traceback (most recent call last):
  File "/Users/edwardlee/git/Chadlia/mujoco-py/src-gen/PandaDemoCamCtrl/PandaDemoCamCtrl.py", line 575, in reaction_function_1
    cv2.imshow("frame", frame.value)
cv2.error: Unknown C++ exception from OpenCV code
*** Terminating app due to uncaught exception 'NSInternalInconsistencyException', reason: 'API misuse: modification of a menu's items on a non-main thread when the menu is part of the main menu. Main menu contents may only be modified from the main thread.'

This error should be avoided by the single-threaded: true target directive, but perhaps there is some incompatibility with mjpython?

[edwardalee]

Again, why not auto-detect?

[edwardalee]

Are you actually using matplotlib? This seems like a lot of requirements.

[edwardalee]

Why not update the existing example rather than copy it here?

[edwardalee]

Again, I don't think we should copy this here.

[edwardalee]

Ah, this may be connected with the error I'm seeing. However, if I uncomment it, it still fails.