Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Initial Analysis of Data-Driven Haptic Search for the Smart Suction Cup

Jungpyo Lee, Sebastian D. Lee, Tae Myung Huh, Hannah S. Stuart

TL;DR

This work evaluates whether a data-driven approach can outperform a prior model-based predictor for determining lateral yaw adjustments ($\phi$) to improve suction-cup grasps with the Smart Suction Cup. It uses a four-sensor pressure array to collect data near a plate edge and compares an MLP-based predictor against a model-based $v_{pred}$-driven method. The results show the data-driven approach yields a lower RMSE ($$17.23 \pm 0.61^\circ$$) than the model-based method ($$20.49 \pm 0.30^\circ$$), suggesting data-driven methods offer advantages even on simple edge geometries while model-based methods remain fast and robust. The authors highlight that data-driven gains depend on representative datasets and propose Sim2Real simulations to streamline integration into more complex grasping scenarios.

Abstract

Suction cups offer a useful gripping solution, particularly in industrial robotics and warehouse applications. Vision-based grasp algorithms, like Dex-Net, show promise but struggle to accurately perceive dark or reflective objects, sub-resolution features, and occlusions, resulting in suction cup grip failures. In our prior work, we designed the Smart Suction Cup, which estimates the flow state within the cup and provides a mechanically resilient end-effector that can inform arm feedback control through a sense of touch. We then demonstrated how this cup's signals enable haptically-driven search behaviors for better grasping points on adversarial objects. This prior work uses a model-based approach to predict the desired motion direction, which opens up the question: does a data-driven approach perform better? This technical report provides an initial analysis harnessing the data previously collected. Specifically, we compare the model-based method with a preliminary data-driven approach to accurately estimate lateral pose adjustment direction for improved grasp success.

Initial Analysis of Data-Driven Haptic Search for the Smart Suction Cup

TL;DR

This work evaluates whether a data-driven approach can outperform a prior model-based predictor for determining lateral yaw adjustments () to improve suction-cup grasps with the Smart Suction Cup. It uses a four-sensor pressure array to collect data near a plate edge and compares an MLP-based predictor against a model-based -driven method. The results show the data-driven approach yields a lower RMSE () than the model-based method (), suggesting data-driven methods offer advantages even on simple edge geometries while model-based methods remain fast and robust. The authors highlight that data-driven gains depend on representative datasets and propose Sim2Real simulations to streamline integration into more complex grasping scenarios.

Abstract

Suction cups offer a useful gripping solution, particularly in industrial robotics and warehouse applications. Vision-based grasp algorithms, like Dex-Net, show promise but struggle to accurately perceive dark or reflective objects, sub-resolution features, and occlusions, resulting in suction cup grip failures. In our prior work, we designed the Smart Suction Cup, which estimates the flow state within the cup and provides a mechanically resilient end-effector that can inform arm feedback control through a sense of touch. We then demonstrated how this cup's signals enable haptically-driven search behaviors for better grasping points on adversarial objects. This prior work uses a model-based approach to predict the desired motion direction, which opens up the question: does a data-driven approach perform better? This technical report provides an initial analysis harnessing the data previously collected. Specifically, we compare the model-based method with a preliminary data-driven approach to accurately estimate lateral pose adjustment direction for improved grasp success.
Paper Structure (7 sections, 1 equation, 2 figures)

This paper contains 7 sections, 1 equation, 2 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Experimental setup and data acquisition
  4. Model-based prediction
  5. Data-driven prediction
  6. Results and Discussion
  7. Conclusion

Figures (2)

  • Figure 1: The schematic drawing of the experimental setup. (a) A lateral offset and the tool frame ($\delta$). (b) Yaw angle ($\phi$) of the suction cup. (c) A lateral direction vector, $v_{pred}$ is predicted to guide motion. Adapted from lee2023haptic.
  • Figure 2: The representative results of yaw angle $\phi$ estimation, showing predicted yaw angle and true yaw angle from data-driven (left) and model-based (right) methods.