TL;DR
This paper introduces a novel 17-degree-of-freedom (DOF) robotic hand designed to perform both dexterous single-object manipulation and complex single-handed assembly. By combining four 4-DOF fingers into two "virtual finger" pairs with a 1-DOF reconfigurable linear palm, the architecture successfully mimics the coordination of two independent hands within a single end-effector.
What Problem Does This Paper Address?
In-hand assembly is a notoriously difficult challenge in robotics. It requires a robot to simultaneously grasp two separate parts and precisely control their relative motion to align, insert, or fasten them.
Traditionally, this is solved using bimanual setups (two robotic arms working together) or specialized end-effectors like orthogonal parallel grippers. However, bimanual setups are bulky and expensive, while specialized grippers lack the dexterity needed to adjust parts locally once grasped. Standard anthropomorphic hands excel at manipulating a single object but struggle to independently manage two objects at once.
The authors bridge this gap by designing a single hand capable of shifting between a unified grasp for large objects and a dual-opposition grasp for two-part assembly tasks.
Methods
The authors developed a 17-DOF reconfigurable hand featuring:
- Four Dexterous Fingers: Based on the LEAP Hand design, each finger possesses 4 DOFs (metacarpophalangeal [MCP] flexion, MCP abduction/adduction, proximal interphalangeal [PIP] flexion, and distal interphalangeal [DIP] flexion).
- A Reconfigurable Palm: A linear actuator provides a 50 mm stroke, allowing two of the fingers (mounted on a prismatic base) to translate relative to the other two (mounted on a fixed base).
Kinematics & Control
The team formulated a kinematic model that mathematically decouples local finger adjustments from the global translation provided by the palm. For single-object manipulation, they leveraged Proximal Policy Optimization (PPO), a reinforcement learning algorithm, to train control policies in simulation that adapt the palm's width based on the target object's size.
Datasets
While this paper does not introduce a standard vision or language dataset, the authors evaluated their system using:
- Simulation Benchmarks: 22 thin spherical shells ranging from 40 mm to 250 mm in diameter, simulated within a physics environment to test grasp workspaces and rotation success.
- Physical Objects: A practice golf ball, a pickleball, and a volleyball were used to validate the simulation-trained reinforcement learning policies on real hardware.
Key Insights and Ideas
- Decoupling Motion Simplifies Control: By using the palm strictly for linear translation (generating assembly forces or adjusting grasp span) and the fingers for local alignment (rotation and fine positioning), the complex problem of two-object manipulation becomes highly manageable.
- The Crucial Role of Abduction: Through ablation studies, the authors discovered that finger abduction/adduction (side-to-side finger movement) is absolutely critical. Without it, the hand cannot sustain the micro-adjustments needed to keep parts aligned during tasks like threading a screw or capping a glue stick.
- Symmetry Breeds Robustness: Distributing four fingers symmetrically into two virtual pairs allows for much more uniform force distribution than standard anthropomorphic layouts, significantly improving the hand's ability to withstand external physical disturbances.
Key Results
The proposed architecture was evaluated against the anthropomorphic LEAP Hand baseline in both simulation and real-world hardware tasks.
Quantitative Performance
- Object Rotation: In simulation, the proposed hand achieved a 98.6% mean success rate for continuous sphere rotation across diameters from 40 to 230 mm, outperforming the LEAP Hand's 73.8%. Notably, for large spheres (210–230 mm), the baseline failed completely, while the reconfigurable hand maintained up to a 100% success rate due to its adjustable palm.
- Disturbance Rejection: When subjected to external force impacts during rotation, a fine-tuned policy on the proposed hand maintained a 92.8% success rate, compared to just 45.2% for the baseline.
Qualitative Hardware Demos
Using teleoperation, the hand successfully executed five distinct in-hand assembly tasks without relying on arm movement to generate relative part motion:
- Through-wall fastening: Holding a nut on one side of a wall while driving and threading a screw from the other.
- Simultaneous screw fastening: Adapting the palm span to drive two separate screws at different spacings.
- Glue-stick capping: Picking up a cap and a base separately, aligning them, and pressing them together.
- Syringe actuation: Stabilizing a syringe barrel with one pair of fingers while the other pair depresses the plunger.
- Cable routing: Passing a cable through an opening using an intra-hand grasp transfer.
Limitations and Future Work
While the mechanical architecture proves highly capable, the assembly tasks presented were entirely teleoperated by a human using a VR headset. The reinforcement learning policies were only applied to single-object rotation.
The authors note that future work will focus on achieving autonomous in-hand assembly by combining imitation learning with visual and tactile feedback. Additionally, they plan to test the architecture on non-spherical objects and explore designs featuring more than two virtual finger oppositions.
Summary updated