Data-Driven Grasp Synthesis - A Survey
TL;DR A survey that sorts data-driven grasp synthesis by what the robot knows about the object beforehand, with stored grasps for known objects, grasps transferred by similarity for familiar ones and heuristics on sensor data for unknown ones. Its tables classify 22, 31 and 16 methods in these three groups.
What they do
A survey of data-driven grasp synthesis: methods that sample candidate grasps and rank them using experience (heuristics, simulation, human demonstration or real robot trials), as opposed to analytic methods that optimize force closure, stability or dexterity. The authors group approaches by what they assume about the object before grasping: known, familiar or unknown.
How it works
- Known objects. Grasps for each object model are generated offline: in simulation on shape primitives or sampled candidates, from human demonstration, or by trial and error. Online, the robot recognizes the object, estimates its pose and filters the stored grasps by reachability.
- Familiar objects. Grasp experience transfers to new objects by similarity: classifiers on 2D, 3D or combined features, matching against a database of exemplars, generative Bayesian networks, or first determining the object’s category.
- Unknown objects. Heuristics map sensor data directly to grasps, by approximating full shape (primitives, symmetry, shape carving), from low-level features, or from global shape.
- Tables I-III tag every method by local or global object-grasp representation, feature type, how grasps are learned, and whether it handles tasks, multi-fingered hands, deformable objects and real data.
Results
- The tables classify 22 approaches for known objects, 31 for familiar and 16 for unknown ones.
- Classic metrics predict real-world success poorly. Grasps stable under them underperformed human grasps taught kinesthetically (Balasubramanian et al.). The epsilon-metric predicted stability poorly under pose error, especially for large objects (Weisz and Allen).
- Data-driven synthesis took off with GraspIt! in 2004; cheap 3D sensing (PR2 projected-texture stereo, Kinect in 2010) then brought many methods working on real 3D data.
- Open problems: most methods assume a pre-segmented object on a planar surface; learning that generalizes from ongoing experience; task constraints; robust execution with tactile and visual feedback.
Limitations
No quantitative comparison: the authors note the lack of common benchmarks, and existing grasp datasets (Columbia Grasp database, VisGraB, playpen) are rarely used for comparison. Task-dependent grasping and the robot hand are only flagged in the tables, not examined. The surveyed data-driven methods give no guarantees of dexterity, equilibrium or stability and can only be verified empirically.
Figures
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.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.