Data-Driven Inverse Kinematics Approximation of a Delta Robot with Stepper Motors

Anni Zhao, Arash Toudeshki, Reza Ehsani, Jian-Qiao SunRobotics 2023 · OpenReview

TL;DR

This paper tackles the challenge of accurately controlling Delta robots, which have complex, nonlinear dynamics. They propose a data-driven approach using neural networks to approximate the robot's inverse kinematics, outperforming traditional geometry-based methods by effectively capturing real-world inaccuracies and achieving superior trajectory tracking.

What Problem Does This Paper Address?

Delta robots are known for their speed and precision, making them useful in tasks like assembly and pick-and-place. However, their "over-actuated" nature and highly nonlinear dynamic models make precise control difficult. A critical component of this control is inverse kinematics, which translates desired end-effector positions into the specific motor angles needed to achieve them. Traditional geometry-based inverse kinematics models often fall short in real-world applications because they struggle to account for manufacturing tolerances, measurement errors, joint flexibility, and other physical imperfections. These inaccuracies lead to significant tracking errors, limiting the robot's performance. The authors aim to bridge this gap by developing a more robust and accurate inverse kinematics model.

Methods

The core of this paper is the development of a data-driven inverse kinematics model for a Delta robot equipped with stepper motors. Instead of relying solely on geometric equations, they train a deep neural network to learn the complex relationship between motor angles and end-effector positions.

Here's a breakdown of their approach:

  • Neural Network Architecture: They employ a deep neural network with 8 hidden layers, each containing 50 neurons, and use a sigmoid activation function.
  • Training Data Generation: Crucially, they generate random experimental data by sampling motor angles within a defined range and then calculating the corresponding end-effector positions. This contrasts with approaches that might use only specific trajectory data.
  • Optimization: The Adam algorithm is used for training, a popular choice for its efficiency in optimizing neural networks.
  • Implementation: The trained neural network is converted into a TensorFlow Lite model for deployment on a Raspberry Pi, enabling real-time control.
  • Open-Loop Control: The validation is performed using an open-loop trajectory tracking setup. This means the neural network directly predicts the next motor command based on the desired trajectory, without feedback from the actual robot position. This method is chosen to directly assess the accuracy of the inverse kinematics model itself.

Datasets

The paper does not introduce a new benchmark dataset. Instead, they generate their own dataset for training and validation:

  • Training Data: Randomly sampled motor angles (converted to joint counts) are used as inputs, and the corresponding end-effector positions (x, y, z) are the outputs. They found that a minimum of 5000 data points is necessary for effective training.
  • Validation Data: The trained model is evaluated on various trajectories (circular, heart, logarithmic spiral) using experimental data collected from the physical Delta robot.

Key Insights and Ideas

  • Random Data is Key for Robustness: The most significant insight is that training neural networks on randomly sampled experimental data leads to a much more robust inverse kinematics model. This approach captures the "rich dynamics" of the robot, including unmodeled factors like manufacturing errors and friction, making the model generalize better to different operating conditions and trajectories.
  • Geometry-Based IK is Fragile: The study highlights that traditional geometry-based inverse kinematics, while theoretically sound, is not robust to real-world uncertainties. This is a critical limitation for practical robotic applications.
  • Open-Loop Validation is Revealing: Using open-loop control for trajectory tracking is an effective way to directly evaluate the accuracy of the inverse kinematics model itself, independent of feedback control loop performance.
  • Data Quantity Matters: The authors emphasize that a sufficient amount of data (at least 5000 points) is crucial for training the neural network to achieve good performance. Insufficient data, especially specific trajectory data, leads to poor generalization.

Key Results

  • Superior Trajectory Tracking: The data-driven neural network model consistently outperforms the geometry-based inverse kinematics in trajectory tracking accuracy across various curves (circular, heart, logarithmic spiral).
  • Reduced Tracking Errors: For circular curves, the neural network achieved tracking errors bounded under 1 cm, whereas the geometry-based method had errors up to 110 cm in the z-axis. Similar significant improvements were observed for other trajectories.
  • Generalization Capability: The neural network model trained on random data demonstrated the ability to generalize well to different trajectories and operating speeds, showcasing its robustness.
  • Flexibility: The neural network model showed excellent flexibility in tracking trajectories at different operating speeds (varying PWM frequencies).

Limitations and Future Work

  • Open-Loop Control: The current validation is performed in an open-loop manner. While this effectively assesses the IK model, a closed-loop system would further minimize tracking errors.
  • Data Generation: The paper relies on generating random data. Future work could explore more sophisticated data generation techniques or leveraging existing datasets if available.
  • Specific Trajectory Data Limitation: The authors explicitly note that neural networks trained on specific trajectory data (rather than random data) "fall short of anticipated performance due to a lack of data." This suggests that while data-driven methods are powerful, the type of data is critical.
  • Future Work: The authors suggest that closed-loop control could further improve performance. They also imply that their data-driven approach can be extended to other robotic systems with complex dynamics.

Summary updated

Abstract

The Delta robot is a parallel robot that is over-actuated and has a highly nonlinear dynamic model, which poses a significant challenge to its control design. The inverse kinematics that maps the motor angles to the position of the end effector is highly nonlinear and extremely important for the control design of the Delta robot. It has been experimentally shown that geometry-based inverse kinematics is not accurate enough to capture the dynamics of the Delta robot due to manufacturing component errors, measurement errors, joint flexibility, backlash, friction, etc. To address this issue, we propose a neural network model to approximate the inverse kinematics of the Delta robot with stepper motors. The neural network model is trained with randomly sampled experimental data and implemented on the hardware in an open-loop control for trajectory tracking. Extensive experimental results show that the neural network model achieves excellent performance in terms of the trajectory tracking of the Delta robot under different operation conditions, and outperforms the geometry-based inverse kinematics model. A critical numerical observation indicates that neural networks trained with the specific trajectory data fall short of anticipated performance due to a lack of data. Conversely, neural networks trained on random experimental data capture the rich dynamics of the Delta robot and are quite robust to model uncertainties compared to geometry-based inverse kinematics.
BibTeX
@inproceedings{zhao2023data,
  author = {Anni Zhao and Arash Toudeshki and Reza Ehsani and Jian-Qiao Sun},
  title = {{Data-Driven Inverse Kinematics Approximation of a Delta Robot with Stepper Motors}},
  booktitle = {Robotics},
  year = {2023},
  url = {https://openreview.net/forum?id=FUF8yZYjR7},
}
Figure