Abstract
Deep learning has significantly advanced 6-DoF grasp detection, yet current models often struggle with "noisy" training labels derived from simplistic physics metrics. Standard force-closure evaluations frequently fail to capture the nuances of real-world stability, leading to unreliable performance in cluttered environments. This paper introduces a Hybrid Physical Metric (HPM) that integrates force-closure with surface flatness, gravity center alignment, and collision safety. By training a novel multi-resolution network, FGC-GraspNet, on these refined labels, we demonstrate a substantial improvement in grasp precision and a 90.5% success rate in real-world robotic manipulation tasks.
Introduction
In robotic manipulation, 6-DoF grasp detection requires a robot to determine the precise 3D position and orientation necessary to pick up an object. Training these models typically involves physics simulators that score millions of potential grasp poses. However, the industry-standard force-closure metric—which determines if an object will slip when squeezed—is often calculated as a binary or coarse discrete score.
This "evaluation insufficiency" creates label noise: two grasps may receive identical scores in a simulator but exhibit vastly different stability in practice. To bridge this gap, researchers at Tsinghua University have developed a more granular scoring system that better reflects the physical realities of stable grasping.
The Hybrid Physical Metric (HPM)
The core innovation of this work is the Hybrid Physical Metric (HPM), a composite score ($S$) that moves beyond simple friction checks. It incorporates four weighted components designed to mimic human grasping intuition:
- Flatness ($S_f$): Quantifies the local surface normals to favor contact with flat areas, which provide more stable friction.
- Gravity Center ($S_g$): Measures the distance between the grasp line and the object's center of mass. Minimizing this distance reduces torque, preventing the object from rotating out of the gripper.
- Collision Perturbation ($S_c$): Evaluates the clearance between the gripper and the object to ensure a safe approach without accidental displacement.
- Force-Closure ($S_t$): Retains the traditional check for basic mechanical stability.
By combining these factors, the HPM provides a continuous and refined label space, reducing gradient noise during model training and allowing the system to distinguish between mediocre and high-quality grasps.
FGC-GraspNet Architecture
To leverage these refined scores, the authors designed FGC-GraspNet, a multi-resolution architecture based on PointNet++. The network is split into two specialized branches:
- Low-Resolution Branch: Focuses on global scene context to segment objects from the background and predict general approach directions.
- High-Resolution Branch: Utilizes a Local Attention Module to analyze fine geometry at contact points. This branch is responsible for the "post" tasks of fine-tuning gripper rotation and depth.
A key feature of the network is its Joint Loss Function, which regresses all potential grasp scores simultaneously. This allows the model to learn the full distribution of grasp quality across the object's surface rather than just identifying a single optimal point.
Experimental Results
The model was trained on the GraspNet-1Billion dataset and validated using a UR3 robotic arm equipped with an OnRobot RG2 gripper.
- Simulation Performance: FGC-GraspNet outperformed the previous state-of-the-art baseline by over 10% in Mean Average Precision (mAP) on seen objects. More importantly, it showed significant gains on novel objects, suggesting that the HPM captures universal physical truths rather than just memorizing training shapes.
- Real-World Success: In complex, cluttered environments—often the most challenging scenario for autonomous robots—the system achieved a 90.5% success rate.
- Ablation Insights: The study revealed that adding a specific depth classification loss improved performance by 7%, highlighting that the depth of the grasp is a critical, yet often undervalued, parameter in 6-DoF detection.
Conclusion
By "mathematizing" human intuition through the Hybrid Physical Metric, this research provides a more robust foundation for training grasp detection models. FGC-GraspNet demonstrates that high-resolution local attention combined with physically grounded scoring can significantly reduce the gap between simulation and real-world performance. Future work will focus on automatically learning the weights for HPM components and adapting the strategy for varying material properties, such as slippery or deformable objects.
Summary updated