Robotic Manipulation Planning for Automatic Peeling of Glass Substrate Based on Online Learning Model Predictive Path Integral

Liwei Hou, Hengsheng Wang, Haoran Zou, Yalin ZhouSensors 2022 · DOI

TL;DR

This paper tackles the challenging problem of automatically peeling glass substrates from adsorption pads, a crucial step in LCD manufacturing. They introduce an "Online Learning Model Predictive Path Integral" (OL-MPPI) algorithm that combines deep learning for system modeling with online adaptation to handle real-world uncertainties, achieving faster and safer peeling than baseline methods.

What Problem Does This Paper Address?

The automated peeling of glass substrates from adsorption pads is a "contact-rich" manipulation task that is notoriously difficult for robots. This process requires precise control to avoid damaging the delicate glass or the pad, and manual labor is currently the standard. Existing robotic solutions struggle with the high degree of nonlinearity and uncertainty inherent in the contact dynamics, such as variations in adhesion, pad properties, and environmental conditions. This paper aims to bridge this gap by developing a robust and adaptive robotic manipulation planning system.

Methods

The core of the proposed solution is the Online Learning Model Predictive Path Integral (OL-MPPI) algorithm. This approach involves several key components:

  • Deep Learning-Based Forward Model: A neural network (combining multi-scale 1D-CNNs for local feature extraction, an attention mechanism, and LSTMs for time-series feature extraction) is used to model the system dynamics. This model predicts the next state of the system based on the current state and control input.
  • Simulation Environment for Pretraining: To overcome the cost and time associated with collecting real-world data, a simulation environment using Pybullet is developed. This environment incorporates dynamic parameter randomization and force compensation to mimic real-world contact behaviors.
  • Offline Learning: The deep learning model is initially trained on data generated from the simulation environment.
  • Online Learning (Meta-Learning): To adapt to real-world uncertainties and avoid catastrophic forgetting, the paper employs a model-agnostic meta-learning (MAML) approach. This allows the model to quickly adapt to new peeling tasks by learning how to learn from a few examples.
  • Model Predictive Path Integral (MPPI): This planning algorithm is used to generate optimal control sequences. MPPI is chosen for its ability to handle nonlinear systems and flexible cost functions, which are essential for this task. The "online learning" aspect means the MPPI planner uses the continuously updated forward model.

Datasets

The paper does not introduce new benchmark datasets. Instead, it relies on:

  • Simulated Trajectories: A large dataset of 50,000 trajectories was generated using a Pybullet simulation environment with randomized dynamic parameters. This dataset was split into training (40,000), validation (5,000), and testing (5,000) sets.
  • Real-World Experimental Data: Data collected during actual glass substrate peeling experiments was used for online learning and evaluating the performance of the OL-MPPI algorithm.

Key Insights and Ideas

  • The Power of Online Adaptation: The most significant insight is the effectiveness of online learning (specifically, meta-learning) in adapting a robot's manipulation strategy to the inherent uncertainties of real-world contact-rich tasks. This allows the system to improve its performance over time as it gathers more experience.
  • Combining Physics and Data: The approach of using a simulation environment with physics-based considerations (dynamic parameter randomization, force compensation) to pretrain a deep learning model is a practical way to leverage prior knowledge and reduce the reliance on extensive real-world data.
  • Multi-Scale Feature Extraction for Contact Dynamics: The use of multi-scale 1D-CNNs and an attention mechanism effectively captures the complex, fluctuating nature of contact forces during the "stick-slip" motion of the peeling process.
  • MPPI's Suitability for Safety-Critical Tasks: The choice of MPPI for planning is justified by its ability to handle complex cost functions that balance task completion speed with safety (minimizing contact forces), making it suitable for delicate operations like glass substrate peeling.

Key Results

  • Improved Performance Over Baselines: The OL-MPPI algorithm significantly outperformed two baseline methods (original MPPI with a static model and a method that used all historical experience for updates) in terms of total cost over 20 consecutive peeling experiments. The total cost for OL-MPPI decreased from 42.22 in the first task to 15.98 by the 11th task.
  • Faster Task Completion: The OL-MPPI algorithm achieved the shortest task completion time (7.1 seconds) compared to Baseline 1 (13.5 seconds) and Baseline 2 (10.4 seconds) in the 20th peeling task.
  • Reduced Wear and Tear: OL-MPPI also resulted in significantly lower total contact forces (∑FRX and ∑FRZ), indicating reduced wear on the blade and adsorption pad. The total contact force for OL-MPPI was 784.64 N for FRX and 486.45 N for FRZ, compared to much higher values for the baselines.
  • Robustness: The system demonstrated robustness, with no glass substrate breakage observed during 20 consecutive peeling experiments.
  • Model Accuracy: The proposed deep learning model achieved better prediction performance compared to MLP and CNN-LSTM models, as indicated by lower RMSE and higher correlation coefficients.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper acknowledges that the OL-MPPI algorithm's performance is influenced by hyperparameters (N, H, and λ). While experimental results are presented, a theoretical basis for hyperparameter selection is still needed.
  • Computational Resources: The current implementation relies on GPU for efficient trajectory sampling, and future work aims to improve efficiency to reduce this dependency.
  • Theoretical Guarantees: The authors plan to theoretically prove the influence mechanism of hyperparameters on the algorithm's performance.
  • Generalization to Other Tasks: While the algorithm is presented as adaptable, further research is needed to explore its application to a wider range of contact-rich manipulation tasks.
  • Mechanism Model Integration: The paper suggests that the forward model could potentially use a pure mechanism model, and the online learning part could correct this mechanism model with a compensation term, which is an avenue for future exploration.

Summary updated

BibTeX
@article{hou2022robotic,
  author = {Liwei Hou and Hengsheng Wang and Haoran Zou and Yalin Zhou},
  title = {{Robotic Manipulation Planning for Automatic Peeling of Glass Substrate Based on Online Learning Model Predictive Path Integral}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22031292},
}
Figure