Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Inverse kinematics learning of a continuum manipulator using limited real time data

Alok Ranjan Sahoo, Pavan Chakraborty

TL;DR

This work tackles data-efficient inverse kinematics for a continuum manipulator under real-time constraints by applying meta-learning. It first trains a MAML-based controller in simulation and adapts it to the real robot with a small number of gradient steps, achieving sub-3% relative error for both known and unknown loading conditions. When a simulator is unavailable, it leverages CGAN-generated data to augment real-world data and again applies MAML, obtaining competitive accuracy though slightly below the simulation-trained approach. The results demonstrate that combining simulator data or CGAN-generated data with fast adaptation can realize accurate, safe, and data-efficient control for continuum manipulators in real time.

Abstract

Data driven control of a continuum manipulator requires a lot of data for training but generating sufficient amount of real time data is not cost efficient. Random actuation of the manipulator can also be unsafe sometimes. Meta learning has been used successfully to adapt to a new environment. Hence, this paper tries to solve the above mentioned problem using meta learning. We consider two cases for that. First, this paper proposes a method to use simulation data for training the model using MAML(Model-Agnostic Meta-Learning). Then, it adapts to the real world using gradient steps. Secondly,if the simulation model is not available or difficult to formulate, then we propose a CGAN(Conditional Generative adversial network)-MAML based method for it. The model is trained using a small amount of real time data and augmented data for different loading conditions. Then, adaptation is done in the real environment. It has been found out from the experiments that the relative positioning error for both the cases are below 3%. The proposed models are experimentally verified on a real continuum manipulator.

Inverse kinematics learning of a continuum manipulator using limited real time data

TL;DR

This work tackles data-efficient inverse kinematics for a continuum manipulator under real-time constraints by applying meta-learning. It first trains a MAML-based controller in simulation and adapts it to the real robot with a small number of gradient steps, achieving sub-3% relative error for both known and unknown loading conditions. When a simulator is unavailable, it leverages CGAN-generated data to augment real-world data and again applies MAML, obtaining competitive accuracy though slightly below the simulation-trained approach. The results demonstrate that combining simulator data or CGAN-generated data with fast adaptation can realize accurate, safe, and data-efficient control for continuum manipulators in real time.

Abstract

Data driven control of a continuum manipulator requires a lot of data for training but generating sufficient amount of real time data is not cost efficient. Random actuation of the manipulator can also be unsafe sometimes. Meta learning has been used successfully to adapt to a new environment. Hence, this paper tries to solve the above mentioned problem using meta learning. We consider two cases for that. First, this paper proposes a method to use simulation data for training the model using MAML(Model-Agnostic Meta-Learning). Then, it adapts to the real world using gradient steps. Secondly,if the simulation model is not available or difficult to formulate, then we propose a CGAN(Conditional Generative adversial network)-MAML based method for it. The model is trained using a small amount of real time data and augmented data for different loading conditions. Then, adaptation is done in the real environment. It has been found out from the experiments that the relative positioning error for both the cases are below 3%. The proposed models are experimentally verified on a real continuum manipulator.
Paper Structure (21 sections, 8 equations, 8 figures, 10 tables, 1 algorithm)

This paper contains 21 sections, 8 equations, 8 figures, 10 tables, 1 algorithm.

Table of Contents

  1. A brief review of MAML
  2. Problem formulation
  3. Architecture of the neural network
  4. Simulation environment
  5. Simulation data collection
  6. Experimental setup
  7. Collecting data from the prototype
  8. Experiments and results
  9. Testing with the single section model
  10. Testing with the double section model
  11. Random point target with different loading condition
  12. Trajectory following experiment with known loading
  13. Trajectory following experiment with unknown external loading
  14. A comparative study
  15. Using GAN for data generation
  16. ...and 6 more sections

Figures (8)

  • Figure 1: Target reaching by the manipulator in real environment
  • Figure 2: Graph of trajectory 1 with known loading (a) 3D view (b) View in ZX plane (c) Tip positioning error. Blue and orange legends show desired trajectory and achieved trajectory respectively
  • Figure 3: Graph of trajectory 2 with known loading (a) 3D view (b) View in ZX plane (c) Tip positioning error. Blue and orange legends show desired trajectory and achieved trajectory respectively
  • Figure 4: Graph of trajectory 1 with unknown loading (a) 3D view (b) View in ZX plane (c) Tip positioning error. Blue and orange legends show desired trajectory and achieved trajectory respectively
  • Figure 5: Graph of trajectory 2 with unknown loading (a) 3D view (b) View in ZX plane (c) Tip positioning error. Blue and orange legends show desired trajectory and achieved trajectory respectively
  • ...and 3 more figures