Geometry, 6D Pose & Vision-Language Robotic Pick-and-Place

A controlled robotics benchmark comparing RGB-D point-cloud grasping, FoundationPose with an oracle initialization mask, and SmolVLA on the same RoboSuite Franka Panda pick-and-place task. Across 240 pilot trials, the point-cloud and FoundationPose systems achieved 92.5% and 96.2% overall success, while a tuned SmolVLA checkpoint reached 72.5%—revealing practical trade-offs between geometric priors, learned visuomotor policies, runtime, and failure modes. The project includes reproducible evaluation tooling, confidence intervals, runtime analysis, and animated demonstrations across four household objects.