Deep Learning a Grasp Function for Grasping under Gripper Pose Uncertainty (1608.02239v1)

Abstract: This paper presents a new method for parallel-jaw grasping of isolated objects from depth images, under large gripper pose uncertainty. Whilst most approaches aim to predict the single best grasp pose from an image, our method first predicts a score for every possible grasp pose, which we denote the grasp function. With this, it is possible to achieve grasping robust to the gripper's pose uncertainty, by smoothing the grasp function with the pose uncertainty function. Therefore, if the single best pose is adjacent to a region of poor grasp quality, that pose will no longer be chosen, and instead a pose will be chosen which is surrounded by a region of high grasp quality. To learn this function, we train a Convolutional Neural Network which takes as input a single depth image of an object, and outputs a score for each grasp pose across the image. Training data for this is generated by use of physics simulation and depth image simulation with 3D object meshes, to enable acquisition of sufficient data without requiring exhaustive real-world experiments. We evaluate with both synthetic and real experiments, and show that the learned grasp score is more robust to gripper pose uncertainty than when this uncertainty is not accounted for.

• The paper introduces a CNN-based approach that predicts a comprehensive grasp function with quality scores to account for gripper pose uncertainty.
• It employs synthetic depth images and physics simulation for large-scale training, resulting in improved grasp success rates in both simulated and real-world experiments.
• Real-world tests using a Kinova MICO arm confirm the method's robustness, demonstrating its potential to enhance robotic grasping in uncertain environments.

Deep Learning a Grasp Function for Grasping under Gripper Pose Uncertainty

This paper presents a neural network-based approach to robotic grasping that aims to address the challenge of gripper pose uncertainty. The authors propose a novel methodology that focuses on predicting a grasp function with quality scores for every possible grasp pose from depth images, rather than identifying a singular optimal grasp pose. This strategy enhances the robustness of robot grasping through smoothing of the grasp function with a pose uncertainty function, thereby avoiding poor grasp choices adjacent to optimal ones.

The approach employs a Convolutional Neural Network (CNN) trained on synthetic depth images generated from 3D object meshes using physics and depth image simulations. This enables the network to be trained on extensive datasets without necessitating exhaustive real-world experiments. Through these simulations, a large and diverse training set is achieved, allowing the CNN to predict a robustness-enhanced grasp function efficiently.

The authors evaluated the effectiveness of their approach through experiments with both synthetic and real-world data. Simulation results demonstrated that the proposed method, dubbed "Robust Best Grasp," consistently outperformed baseline strategies, particularly under higher pose uncertainties. For instance, the method exhibited notable improvements in grasp success rates when compared to strategies that did not account for pose uncertainty, even under substantial deviations in gripper positioning.

In real-world experiments using a Kinova MICO robotic arm equipped with a depth camera, the proposed method also showed superior performance. The results indicated adaptability to real-world scenarios despite the model being trained on synthetic datasets. This suggests that the simulated training environment was instrumental in effectively capturing essential dynamics for real-world applications.

The main contribution of this paper lies in its innovative use of a CNN to learn a comprehensive grasp function that integrates uncertainty considerations, improving the dependability of robotic grasping. This work provides a foundation for future advancements in robotics, particularly in developing systems that can operate effectively amid significant uncertainties in control and sensory feedback.

Furthermore, this research opens paths for several future developments, such as examining the incorporation of dynamic uncertainties and exploring grasp strategies involving multi-fingered hands or more complex robotic grippers. Exploring active learning methods to optimize the selection of simulations could also enhance the scalability of training frameworks, allowing for even more extensive and diverse training datasets.

Overall, this paper offers a robust framework for enhancing robotic grasping via deep learning, with significant implications for practical implementations in areas where precision in grasping under uncertainty is crucial, such as service robotics and automated manufacturing.