Introduction
Robotic grasping in unstructured, cluttered environments remains a significant challenge, particularly when dealing with novel objects. Traditional model-based methods rely on pre-existing 3D CAD models, while model-free approaches often require vast datasets and struggle to generalize across different object scales. SymmetryGrasp addresses these limitations by introducing a learning-based framework that utilizes object symmetry as a geometric prior. By explicitly estimating symmetry from single-view RGB-D images, the system can "fill in the blanks" of occluded geometries, enabling more reliable grasp detection.
Methodology
The SymmetryGrasp pipeline is a multi-stage architecture designed to move from raw visual data to precise robotic control.
- Symmetric Region Detection: The process begins with a Mask R-CNN that identifies graspable, symmetric regions within the RGB-D input.
- Symmetry-Aware Grasp Network: Cropped point clouds are processed through a PointNet++ backbone. This network is uniquely trained to simultaneously predict grasp affordance (the likelihood of a successful contact point) and symmetry parameters (planes or axes of symmetry).
- Principal-directional scale-Invariant Feature Transformer (PIFT): A core innovation of this work, PIFT is a module that ensures feature invariance across different object scales along principal axes. This allows the model to recognize similar grasp geometries in objects of varying sizes, such as different types of bottles or containers.
- Two-Stage Parameterization: Rather than predicting a full 7-DoF grasp directly, the model identifies two symmetric contact points to establish 6-DoF and gripper width, followed by a final rotation prediction around the contact axis.
Key Innovations and Insights
Symmetry as a Geometric Prior
By forcing the neural network to estimate symmetry parameters, the model develops a more robust internal representation of 3D geometry. This symmetry branch acts as a regularizer, narrowing the search space for feasible grasps and allowing the robot to propose contact points on the "dark side" of objects that are occluded from the camera's single viewpoint.
Scale Invariance via PIFT
Standard point cloud networks often fail when objects are anisotropically scaled (stretched or squashed). PIFT addresses this by normalizing features along the object's principal directions. To reinforce this, the authors employ a Siamese training architecture to ensure feature consistency between original and scaled versions of the same object.
Experimental Results
The framework was evaluated against leading benchmarks and on physical hardware, demonstrating superior performance in complex scenarios.
- GraspNet-1-Billion Benchmark: SymmetryGrasp achieved a Grasp Quality score of 0.382 on seen objects and 0.297 on unseen objects, outperforming previous state-of-the-art methods (0.291 and 0.276, respectively).
- Stability: The system reached an Antipodal Score of 0.809, indicating a high frequency of stable, force-closure grasps.
- Real-World Validation: In tests using a UR-5 robot arm, the system achieved a 74% success rate for objects with reflectional symmetry, proving its resilience to the noise and calibration errors inherent in physical hardware.
Conclusion and Future Work
SymmetryGrasp demonstrates that incorporating geometric intuition—specifically symmetry—significantly enhances a robot's ability to interact with novel objects in cluttered scenes. While the system excels with symmetric man-made objects, performance decreases on irregular organic shapes (47% success rate). Future research will explore equivariant networks to further improve geometric reasoning and investigate the application of symmetry priors in more complex manipulation tasks like robotic assembly.
Summary updated