Standing Peg-in-Hole Insertion: Demonstrations, Trained Policies, Evaluation Episodes, and Simulation Environment

Reproducibility bundle for the manuscript "Perception or Policy: Sub-Task Failure Attribution in Sub-Millimeter-Clearance Robotic Assembly". A UR5e + Robotiq 2F-85 grasps a standing pipe (20 mm diameter), transports it over a holder, and inserts it into a bore (21.5 mm diameter, 0.75 mm radial clearance) in an Isaac Sim cell with three calibrated cameras (wrist D405, side D455, overhead Kinect; 256x256 RGB). Data generation is domain-randomized; evaluation fixes the hardest configuration in that distribution - the 21.5 mm holder at nominal 0.75 mm clearance, bore-fit randomization off. The record contains: (1) the LeRobot demonstration dataset (1565 episodes / 197,190 frames; 10-D absolute-pose action, RGB and RGB-D variants); (2) all trained models - a ResNet-18 grasp-pose regressor and three fine-tuned end-to-end policies (Diffusion Policy 200k, SmolVLA 40k, pi0.5 30k); (3) the held-out standing evaluation episodes (N=50, indices 2000-2049, seed 42) behind the paper's tables, including the handoff-height ablation logs; and (4) the perception package plus the simulation environment and evaluation harness, with a standalone no-Isaac inference demo. Grading: the bundled grader reports two criteria. "completion" (the default, and the one the manuscript reports) asks whether the robot picked the pipe and placed it upright at the holder: lateral < 25 mm, tilt < 15 deg, lift > 50 mm. "insertion" adds a height gate (< 30 mm above the seat) and asks whether the pipe entered the bore. The two differ on seven episodes, all in the learned-policy runs; the scripted and privileged rows are identical under both. README.md reports both columns. See README.md for the mapping of each evaluation set to the paper's results and for loading/reproduction instructions.