Underactuated Gripper Results: Sim-to-Real Verification

Table of Links

Abstract and I. Introduction

II. Methods

III. Generation of Tendon Driven Grippers

IV. Results and Discussion

V. Conclusion and References

IV. RESULTS DISCUSSION

A. Search results

The algorithm evaluates and saves encoded mechanism representations and rewards during the design search process. We consider the designs with top rewards as a set of final results. The designs generated are shown in Fig. 5. The quantitative characteristics of the MCTS run are shown in Fig. 9. The increasing V and Q functions demonstrate that the search algorithm consistently achieves higher average rewards over time.

The search space expands to include more valuable designs. A test run, that is performed in each step illustrates the current learned model of the search space. After 40 iterations, the reward increases significantly. Using the learned design space model, the MCTS constructed a decision tree that extended to terminal states. It then proceeded to choose rules based on this tree. Once the learned MCTS has gone through a significant number of iterations, it constructs a decision tree that leads to terminal states with a highly promising outcome. Simulation took 2 days on a machine with CPU i9 10920X with 12 cores to conduct all calculations.

B. Experimental setup

We created a physical setup to validate our approach and verify the results (Fig. 8). The CAD model was designed with consideration of the graph grammar described in IIIB, which ensures that any generated graph could be tested in the physical world in a fast and efficient manner. The phalanx was chosen as the foundational building block of the prototype, and it was designed to be adjustable. Each link is equipped with a prismatic joint (Fig. 4), such that its length can be varied. This versatility allowed us to assemble various configurations with different numbers of fingers, number of phalanxes and lengths. The design includes features from the tendon-driven mechanism such as pulleys for directing the tendons and springs. The position of the pulleys and the stiffness of the springs were selected to resemble the virtual experiments for more credible validation of our methodology.

Additionally, the palm component was specifically designed to facilitate testing across all finger positions and orientations that our method can produce. In terms of manufacturing, we employed 3D printing for the phalanges and laser cutting for the palm for The Dynamixel AX-12A servo were chosen to control the fingers due to their suitable torque characteristics and diverse feedback capabilities. We added FSR402 sensors to measure grasping pressure. To manage and communicate with these sensors and motors, we used an STM32F4 microcontroller

C. Verification

The tendon-driven finger is an underactuated open chain mechanism, thus its movement pattern and contact forces depend on both tendon tension and pulley positions. In order to verify the precision of the testing setup, we carried out a set of experiments with a singular finger. The testing setup is illustrated in Fig. 7. Motion of the physical finger has been compared with the virtual one by visual inspection. We measured the contact force between the test finger and the fixed object using the FSR402 sensor to determine the difference between the simulation and the experiment. The difference in simulated and experimental contact forces in our tests did not exceed 34% (Fig. 7).

It was noticed, that a larger number of fingers have been generated on the side where the external force has been applied. The grasp is more secured if the fingers wrap around from different directions like tentacles. Thus, as for the grippers on Fig. 8 we have analyzed the grasping patters visually. Because of underactuation, a significant effect on object initial state, and ”flying” objects the motion of the fingers differs from the simulation. Nevertheless, because we have trained them on a set of objects with randomly applied forces, the grippers are versatile enough to encompass the objects even if they are oriented differently.