A Generalized Placeability Metric for Model-Free Unified Pick-and-Place Reasoning

TL;DR

The paper addresses robust pick-and-place for unknown objects under sensor noise by introducing a model-free, CAD-free generalized placeability metric that operates directly on partial point clouds. It fuses stability, altitude-based clearance, and placement-conditioned graspability to evaluate 6-DoF placements and enable unified pick-and-place reasoning without CAD priors. Key contributions include the object-centric formulation, three-component placeability metric with explicit mathematical definitions, and a unified scoring framework that links grasp feasibility to downstream placements. The approach demonstrates improved stability prediction, edge-and-incline tipping handling, and real-world applicability on a UR5 robot in cluttered shelves, offering a scalable and online alternative to prior CAD- or plane-dependent methods with practical impact for robotics manipulation in unstructured environments.

Abstract

To reliably pick and place unknown objects under real-world sensing noise remains a challenging task, as existing methods rely on strong object priors (e.g., CAD models), or planar-support assumptions, limiting generalization and unified reasoning between grasping and placing. In this work, we introduce a generalized placeability metric that evaluates placement poses directly from noisy point clouds, without any shape priors. The metric jointly scores stability, graspability, and clearance. From raw geometry, we extract the support surfaces of the object to generate diverse candidates for multi-orientation placement and sample contacts that satisfy collision and stability constraints. By conditioning grasp scores on each candidate placement, our proposed method enables model-free unified pick-and-place reasoning and selects grasp-place pairs that lead to stable, collision-free placements. On unseen real objects and non-planar object supports, our metric delivers CAD-comparable accuracy in predicting stability loss and generally produces more physically plausible placements than learning-based predictors.

Figures (10)

• Figure 1: From a noisy object point cloud, our system evaluates candidate placements using the proposed generalized placeability metric, which scores stability, graspability, and clearance directly on sensor data. This metric enables model-free unified pick-and-place reasoning: by jointly scoring grasp and placement candidates, the system selects admissible grasps that lead to the most stable placements in cluttered real-world scenes.
• Figure 2: Overview of our framework for model-free unified pick-and-place reasoning. The pipeline integrates perception, grasp generation, and placement evaluation into a single object-centric process. Point clouds are reconstructed from RGB-D observations to generate object and environment representations. Feasible grasps and candidate placements are then evaluated using our generalized placeability metric, which accounts for stability, placement-conditioned graspability (PCG), and clearance. A unified score combines these factors with heuristic constraints to select the best-scoring pick-and-place pair. Finally, motion planning and execution proceed by iterating through candidate pairs in descending score order until a feasible trajectory is found.
• Figure 3: Stability measure, illustrated on a noisy partial point cloud of a power drill with its front section missing. (a) Stable: most sampled points lie inside the support polygon, yielding a high score. (b) Unstable: most points fall outside, giving a near-zero score. (c) Inclined: despite full contact, the steep tilt projects points beyond the polygon, again producing a low score.
• Figure 4: Robotic setup used for the experiments.
• Figure 5: Qualitative comparison of predicted stable surfaces between our method and UOP noh2024learning. Our approach consistently identifies a physically stable placement even from noisy partial point cloud data, whereas UOP often predicts unstable placements. Both methods place the mustard bottle on stable contact points. In the case of the power drill, it cannot be placed on the surface predicted by UOP due to its curved geometry, whereas our method correctly identifies the peak contact points that allow for a stable placement.