Unitree G1's VLA requires 4 RealSense cameras, devs seek alternatives
Vision-language-action models need 2 wrist + 2 shoulder cameras per robot.
A new challenge has emerged in the humanoid robotics community: deploying Vision-Language-Action (VLA) models for continuous tasks on robots like Unitree's G1 and EngineAI's T800. According to a developer post on the ROS General forum, Unitree's open-source unifolm VLA (hosted on GitHub) requires a specific camera setup—two Intel RealSense depth cameras on the wrists and two more on the shoulders. This four-camera configuration is presented as a hard requirement, and the developer admits being "kinda stuck" on whether there is any way around using RealSense hardware.
The post sparked discussion about the broader dependency on Intel RealSense in VLA pipelines. While RealSense cameras are popular for their depth sensing and ROS integration, the rigid requirement limits flexibility and adds cost to already expensive humanoid platforms. Community members are exploring alternatives such as Orbbec depth cameras, stereo vision pairs, or simulated data augmentation, but no turnkey solution has emerged yet. This bottleneck matters because VLA models are key to teaching robots long-horizon tasks—like assembly or kitchen work—and hardware constraints could slow real-world adoption.
- Unitree's unifolm VLA mandates 4 Intel RealSense cameras (2 wrist, 2 shoulder) on humanoids like G1 and T800.
- Developer on ROS General forum reports being blocked without an alternative to RealSense for VLA deployment.
- No standard workaround yet; community weighs Orbbec, stereo vision, and simulation as potential substitutes.
Why It Matters
Camera hardware constraints could bottleneck VLA deployment on humanoids, raising costs and slowing real-world robotics adoption.