A new open-source runtime called vla.cpp lets eleven different robot AI models share a single C++ inference engine instead of eleven separate ones.
Researchers built vla.cpp to strip PyTorch out of the loop for running vision-language-action models, the systems that turn camera feeds and text instructions into robot motor commands. The runtime shares core plumbing - model loading, tensor execution, serving - across all eleven supported architectures, while letting each model keep its own attention, conditioning, and action-generation logic. On the LIBERO-Object benchmark, one model, BitVLA, completed all 200 test episodes on an 8GB Jetson Orin Nano, a modest onboard computer rather than a full workstation GPU. A specialized ternary tensor-core kernel also sped up BitVLA's inference by 4.0 to 4.6 times over a standard CUDA-core baseline on RTX 3060 and AGX Orin hardware.
Robotics teams currently rewrite inference code for every new model-and-chip pairing, which slows testing and real-world rollout. A shared runtime that still hits a perfect task-completion score on cheap onboard hardware suggests robot AI doesn't need a workstation GPU to work reliably. A separate case study in the paper also found that positional-index precision, not just model choice, can quietly break gripper commands - a reminder that benchmark scores alone don't catch every failure mode.
The name is no accident: like llama.cpp did for language models, vla.cpp is betting that a lean, dependency-free runtime matters as much as any individual model's benchmark score.