Abstract
This paper introduces the Dexterous Manipulation Graph (DMG), a framework enabling robots with simple parallel grippers to perform complex in-hand object repositioning. By modeling finger movements as a graph and utilizing dual-arm coordination, the system allows robots to transition between grasps through a sequence of slides, rotations, and hand-overs, effectively overcoming the dexterity limitations of standard two-fingered grippers.
Introduction
Most industrial and service robots utilize simple parallel grippers that lack the inherent dexterity of a human hand. To change a grasp, these robots typically must perform a "regrasp" by placing the object on a flat surface and picking it up again—a slow process that requires specific environmental conditions. While some methods use "extrinsic dexterity" (leveraging gravity or external surfaces), planning these motions for complex shapes is mathematically challenging. This work addresses these issues by planning a seamless sequence of in-hand slides, rotations, and hand-overs to move from an initial grasp to a target configuration without releasing the object.
Methodology
The authors propose a three-step pipeline to achieve dexterous repositioning: DMG generation, planning, and execution.
1. DMG Generation
The object’s surface is subdivided into small areas called "supervoxels." The system constructs a graph where each node represents a possible contact point and a range of collision-free finger orientations. Edges are created between nodes if a finger can move between them using specific primitives.
2. Planning
The problem is treated as a graph search using Dijkstra’s algorithm to find the shortest path of primitives from the start to the goal. A key innovation is the handling of graph discontinuities. If a finger cannot physically slide across a certain feature (like a sharp edge), the system plans a regrasp—using the second robot arm to take the object so the first arm can reposition itself on the graph.
3. Execution
The plan is executed by a dual-arm ABB YuMi robot. The system employs Extended Cooperative Task Space (ECTS) control, allowing the arms to share the task: one arm maintains a controlled grip while the other pushes the object to facilitate the planned sliding or pivoting motion.
Key Insights and Results
- The Graph as a Roadmap: Pre-computing the DMG transforms a complex physics problem into a standard pathfinding problem. To simplify the search space, the system plans for one "principal" finger and verifies that a "secondary" finger can follow on the opposite side.
- Dual-Arm Synergy: The second arm acts as a dynamic support, making the manipulation "extrinsic" to the hand but "intrinsic" to the robot system, offering significantly higher flexibility than static environments.
- Performance: Generating the DMG for complex objects takes between 1.8 and 11.6 seconds on a standard CPU, after which pathfinding is nearly instantaneous.
- Validation: The system was successfully tested on synthetic models (including an L-shaped peg with concavities) and real-world objects from the YCB dataset, such as a mug. The robot demonstrated the ability to hand off objects between arms to navigate around geometric obstacles.
Limitations and Future Work
Currently, the system requires a known 3D model of the object. Future research aims to generate DMGs on-the-fly from raw sensor data, such as depth cameras. Additionally, the authors plan to integrate robot kinematic constraints (joint limits) directly into the graph costs and explore scaling the DMG concept to multi-fingered dexterous hands.
Summary updated