CLAMP: A Sim-to-Real Benchmark for Closed-Loop Kinematic Pose Estimation and Assembly Reasoning

Existing benchmarks for known-object pose estimation either treat objects as rigid bodies or restrict articulation to serial-chain robot arms with revolute joints, and all assume that every part is always present. Real-world industrial and consumer equipment, however, exhibits closed-loop kinematic chains, prismatic joints, and structural assembly perturbations—challenges that current datasets and methods do not address. We introduce CLAMP (Closed-Loop Assembly and Mechanism Perception), a sim-to-real benchmark for kinematic pose estimation and assembly reasoning on mechanically complex known objects. The dataset spans seven diverse pieces of equipment—ranging from a delta 3D printer with 27 coupled DOFs to a hydraulic jack with closed kinematic loops—comprising over two million synthetic training images and approximately 21,000 real test images across 210 scenes, each annotated with ground-truth joint states and part-level assembly labels. To produce this data we build supporting infrastructure, released with the benchmark: (i) a graph-based kinematic representation that extends URDF's tree structure with closed-loop constraints and the topology changes induced by missing parts; (ii) a synthetic data generator that samples valid poses on the kinematic constraint manifold and stochastically removes parts with automatic kinematic restructuring; and (iii) a real-image labeling pipeline that jointly recovers camera alignment and equipment pose via multi-view constrained optimization. The pipeline requires only a CAD model and kinematic description for each new piece of equipment, with no per-equipment code, and the seven types serve as representative test cases of real-world mechanical complexity. We adapt three methods from different families as baselines: RoboPEPP, the state of the art on the DREAM robot pose benchmark, as well as HoRoPose and CRAVES. An ablation shows that our constrained pose sampling improves accuracy for most closed-loop equipment. The best method varies by equipment, but none performs consistently well: even the strongest on average, RoboPEPP, remains well below its accuracy on DREAM, and all three degrade on average when parts are missing. We release the full dataset (https://huggingface.co/datasets/clamp-benchmark/clamp-benchmark) as well as the synthetic data generation pipeline, evaluation code, and baseline implementations (https://github.com/overlab-kevin/clamp).