In traditional model-based control methods, the controller directly reasons with the dynamic model of the robot. Recent studies use policies established through reinforcement learning as robotic structures become more complex and biomimetic. This is especially true for operations requiring skill, such as manipulation involving numerous fingers and the application of an anthropomorphic robotic hand. The ability to move in unison can revolutionize several industries, from pick-and-place warehouse work to assembly line manufacturing to assist in the home.
New research by ETH Zurich and Max Planck ETH Center for Learning Systems introduces the Faive Hand as a dexterous manipulation platform. As a first step toward humanlike manipulation, the team reports on their current work integrating its model into an RL environment and applying a closed-loop controller on the robot to achieve dexterous in-hand spherical rotation.
The most prominent robotic hands are currently employed for dexterous manipulation research, considering that capable robots require both hardware and a controller. The researchers propose that a more humanlike hand design is more suited for engaging with tools and items in the environment because they were made with people in mind from the beginning. When learning from human examples, manipulation activities are easier to transmit to a robot with a similar framework.
The Faive Hand was created in the Soft Robotics Lab as a biomimetic, tendon-driven robotic platform for investigating fine-grained manipulation. The newest iteration of the hand is 3D printed and powered by servo motors, making mass production easy and accessible. However, unlike other dexterous hands taught with RL, this hand incorporates features such as rolling contact joints that rotate without a defined axis of rotation, adding difficulty to the already difficult task of controlling a high-DoF robotic hand for manipulation. Since conventional rotational encoders are challenging to implement in this design, internal joint angle encoders are still in the works but must be included in the hand. Due to this restriction, the servo motor angles are used to estimate the tendon length and, hence, the joint angles. With these additions to the simulation framework and the low-level controller, a policy trained with closed-loop RL may be executed on the real robot.
The researchers demonstrate the hand’s potential by demonstrating a zero-shot transfer of skills taught with RL in the IsaacGym simulator. They plan to improve it at RL sim2real tasks and other tasks like behavior cloning by adding actuation and sensor capabilities.