New ML method helps mobile robots quickly find optimal positions to grab objects

Researchers have developed GBPP (Grasp-Aware Base Placement Prediction), a machine learning system that helps mobile robots decide where to position themselves before attempting to grasp an object. Rather than relying on complex motion-planning calculations, the method evaluates candidate poses in real time using a learned scoring model.

The Two-Stage Training Approach#

The system uses a practical curriculum learning strategy to build an accurate predictor with limited labeled data. In the first stage, a simple rule based on distance and visibility constraints auto-labels a large dataset quickly and cheaply, providing broad coverage of the pose space. The second stage then refines the model using a smaller set of higher-fidelity simulation trials that capture true grasp success outcomes. This pairing of inexpensive heuristics followed by targeted simulation refinement makes the training process data efficient and scalable.

How It Works#

The method encodes a single RGB-D image using a PointNet++ style point cloud neural network, which feeds into a multilayer perceptron (MLP) scorer. This architecture evaluates dense grids of candidate base poses without requiring full task-and-motion planning—enabling rapid online selection in real robot scenarios. The resulting predictions are faster to compute while maintaining awareness of the 3D geometry present in the scene.

Real-World Performance#

Testing on both simulated environments and a physical mobile manipulator showed that GBPP outperforms simpler baselines that rely only on proximity or geometric cues. The model selects more stable stances, avoids unreachable configurations, and degrades gracefully when its predictions are incorrect. The approach offers a practical recipe for geometry-aware robot base placement that balances computational speed with grasp success.

View full paper on arXiv