Articulated Object Manipulation

The field asks robots to infer an object's joints, movable parts, and current state, then open, close, or dexterously use it despite uncertain geometry and contact. Approaches range from RGB-D and point-cloud kinematic estimation and semantic scene graphs to constraint-aware control, imitation learning, reinforcement learning, and vision-language-action policies trained on synthetic or human demonstrations. Evaluation spans simulated suites such as PartNet-Mobility and PartManip, as well as real doors, drawers, appliances, and tools, with growing attention to mobile manipulation and transfer across object categories and robot embodiments. A central tension is that policies must generalize from limited data while respecting real joint mechanics and contact forces; simulation fidelity, hidden kinematics, and reliable sim-to-real transfer remain difficult.