RigidBench: simulator-grounded evaluation data for rigid-body physics in video generation

RigidBench is a 100-sample evaluation set for measuring rigid-body physics in image-to-video models. Each sample pairs a photorealistic Blender Cycles reference video with simulator ground truth, including per-actor masks, per-pixel depth, 6-DoF trajectories, and contact events. The dataset covers five tasks: free fall, bounce, ramp collision, ball chain, and drop into cluster. The four evaluation partitions vary whether scenes and objects appear in the simulator-generated training set. The archive includes a manifest, prompts, sample metadata, and loading documentation.