So, current issue of the launching VLAs on the Humanoids, like G1 Unitree\EngineAI T800\another,
is the realsense pack of cams. F.e., for the unifolm vla (GitHub - unitreerobotics/unifolm-vla · GitHub), you need 2 on wrists + 2 on shoulders, so its kinda there is no another way without realsense?
Currently kinda stuck on this type of the issue, maybe someone know the solution?
On the camera side, one thing that bit me and is not obvious.
If you are running a VLA on a humanoid for continuous tasks, the input pipeline is part of your attack surface, not just your perception stack. I run adversarial evaluation on VLA policies, and the cheapest successful attacks in my suite are not clever prompts; they are things that change what the policy sees or how the instruction is framed.
Not saying that is your problem here. Saying that if you get the camera working and the policy still behaves oddly on continuous tasks, look at the instruction path before you look at the weights.