Hierarchical Tactile-Based Control Decomposition of Dexterous In-Hand Manipulation Tasks | JoVE Visualize

Area of Science:

Robotics and Control Systems
Artificial Intelligence
Sensor Fusion

Background:

In-hand manipulation and grasp adjustment in dexterous robotic hands are challenging due to complex finger coordination and interaction variability.
Integrating tactile feedback into robotic control significantly increases complexity, with traditional methods often lacking this capability or relying on overly simplified models.
Existing approaches struggle with unconstrained manipulation tasks, often requiring problem simplification or complex, unavailable models.

Purpose of the Study:

To propose a novel hierarchical control approach for dexterous robotic hands that incorporates tactile feedback.
To enable reinforcement learning (RL) methods to learn complex, unconstrained manipulation tasks by structuring the state-action space.
To develop low-level controllers that ensure grip stability using tactile information, providing an abstraction for sensor input.

Main Methods:

A hierarchical control architecture is proposed, combining a high-level policy learned via reinforcement learning with low-level grip stabilization controllers.
Low-level controllers utilize tactile feedback independently to maintain grip stability during manipulation.
The structured exploration facilitated by the low-level controllers allows RL agents to learn complex tasks.

Main Results:

The hierarchical structure enables RL methods to learn unconstrained manipulation tasks that are intractable in non-hierarchical settings.
Low-level tactile controllers provide an abstraction layer, simplifying sensor input for the RL policy.
Preliminary results demonstrate successful transfer of policies trained in simulation to a real robot hand platform.

Conclusions:

The proposed hierarchical control approach effectively addresses the complexities of tactile-informed in-hand manipulation for robotic hands.
This method enhances the capabilities of reinforcement learning in robotics, allowing for learning of more complex manipulation skills.
The developed system shows promise for real-world robotic applications through successful sim-to-real transfer.