Data-Driven Grasp Synthesis - A Survey (1309.2660v2)

Abstract: We review the work on data-driven grasp synthesis and the methodologies for sampling and ranking candidate grasps. We divide the approaches into three groups based on whether they synthesize grasps for known, familiar or unknown objects. This structure allows us to identify common object representations and perceptual processes that facilitate the employed data-driven grasp synthesis technique. In the case of known objects, we concentrate on the approaches that are based on object recognition and pose estimation. In the case of familiar objects, the techniques use some form of a similarity matching to a set of previously encountered objects. Finally for the approaches dealing with unknown objects, the core part is the extraction of specific features that are indicative of good grasps. Our survey provides an overview of the different methodologies and discusses open problems in the area of robot grasping. We also draw a parallel to the classical approaches that rely on analytic formulations.

• The paper presents a comprehensive review that categorizes robotic grasp synthesis into known, familiar, and unknown object approaches.
• It compares analytical methods with data-driven techniques, emphasizing empirical data, object recognition, and sensory feedback to improve grasp performance.
• The survey highlights future research directions, including adaptive learning and multisensory integration for more robust autonomous manipulation.

Data-Driven Grasp Synthesis: A Survey

The paper "Data-Driven Grasp Synthesis - A Survey" by Jeannette Bohg, Antonio Morales, Tamim Asfour, and Danica Kragic, published in the Transactions on Robotics, provides a comprehensive review of methodologies for sampling and ranking candidate grasps for robotic grasp synthesis. The survey classifies these methodologies into three primary categories based on the level of prior knowledge about the objects: known, familiar, and unknown objects.

Analytical Methodologies

The authors begin with a brief overview of classical analytical approaches which use geometric, kinematic, or dynamic formulations to guarantee specific grasp metrics. These approaches typically configure force-closure grasps with multi-fingered robotic hands. Despite their theoretical guarantees, they often rely on simplifications such as rigid body modeling, Coulomb friction, and precise geometric and physical models, which limit their practical applicability. For instance, Prattichizzo and Trinkle highlight the limitations of these models in accounting for real-world inconsistencies, such as dynamic hand-object interactions and actuator inaccuracies due to sensor noise.

Data-Driven Approaches

In contrast, data-driven methodologies, which form the crux of this survey, primarily rely on empirical data and heuristics derived from human demonstrations or autonomous trial and error during the execution on robotic platforms. These methods generally fall into the following categories:

1. Known Objects: These methods assume the object has been previously modeled, allowing the use of object recognition and pose estimation. Once recognized, the grasp configurations stored in a database can be recalled and applied. Examples include the GraspIt! framework and simulations by Goldfeder et al. that utilize superquadric approximations for object modeling.
2. Familiar Objects: For objects that have not been encountered but are similar to known objects, grasp synthesis relies on feature similarities. Techniques involve object categorization and transfer of grasp experiences from known objects. Methods by El-Khoury and Sahbani leverage object part segmentation to determine graspable areas heuristically, while Saxena et al. and Bohg et al. use machine learning to predict grasp points based on visual features.
3. Unknown Objects: Synthesis for unknown objects pertains to identifying structural features directly from sensory data to generate and rank grasps. Approaches like shape fitting techniques used by Dunes et al., Rao et al., and Bohg et al. employ symmetry and local geometry for heuristic grasp generation.

Evaluation and Robustness

While numerous methods, including Bicchi and Kumar's and related studies, show that grasps synthesized by classical metrics may perform suboptimally in real-world conditions, data-driven approaches are generally validated through simulation and empirical tests. However, they often need to adapt to perturbations and uncertainties inherent in real-world scenarios. The survey indicates a trend towards integrating tactile and visual feedback for more robust execution, as demonstrated by Hsiao et al. using proximity sensors and online reachability filtering.

Implications and Future Developments

The implications of data-driven grasp synthesis methodologies are significant for both theoretical and practical aspects of robotics. Theoretically, they push the boundaries of how unknown environments can be navigated and manipulated autonomously. Practically, these methodologies aim to enhance robotic performance in unstructured and unforeseen conditions, enabling applications such as household robotics, manufacturing automation, and service robots.

Future research, as inferred from the survey, should focus on:

• Developing more universal and adaptive perceptual processing techniques for robust object segmentation and feature extraction.
• Exploring life-long learning systems that allow robots to continuously adapt to new objects and scenarios based on accumulated experiences.
• Integrating multisensory feedback mechanisms to enhance the reliability and success rate of grasps in uncertain conditions.

In summary, the surveyed paper illustrates a significant evolution from classical to data-driven approaches in robotic grasp synthesis. The comparative analysis of various methodologies highlights their strengths and exposes gaps that could guide future research endeavors in advancing autonomous robotic manipulation systems.