From Automation to Autonomy: Deploying Vision‑Language‑Action Models on Robotic Manipulators

Deploying vision-language and vision-language-action models on constrained robotic platforms requires evidence that their spatial outputs and learned actions remain useful with limited compute and data. Two complementary systems are investigated. First, a locally hosted vision-language model produces pixel-space targets that a planar homography maps to commands for a Mitsubishi industrial manipulator. Second, ACT, Diffusion Policy, SmolVLA, and $\pi_{0.5}$ are trained on 80 kinesthetic demonstrations comprising 9{,}193 image-state frames from three cameras and deployed on a desktop myCobot 280 arm. Released PushT and LIBERO checkpoints first validate the shared evaluation path. In a 20-frame synthetic localization benchmark with manually checked model-generated reference coordinates, Florence-2-large and Grounding-DINO-tiny return valid outputs for every frame and place 18 and 16 predictions within tolerance. On the physical pick-and-place task, SmolVLA and $\pi_{0.5}$ each complete 14 of 20 trials, while ACT completes 5 and Diffusion Policy 3; stage-wise analysis locates the principal separation at grasping rather than reaching. Although $\pi_{0.5}$ achieves the highest LIBERO scores, it does not fit the available 8\,GB GPU and therefore runs through CPU inference during robot deployment. Under this constraint, it matches the physical placement count of the 450\,M-parameter SmolVLA. The study establishes a reproducible route from local model evaluation to physical deployment and identifies hardware compatibility as central to policy selection for a real robotic system.