Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reinforcement Learning-Based Model Matching to Reduce the Sim-Real Gap in COBRA

Adarsh Salagame, Harin Kumar Nallaguntla, Bardia Ardakanian, Eric Sihite, Gunar Schirner, Alireza Ramezani

TL;DR

A reinforcement learning-based model identification method aimed at enhancing the accuracy of the dynamics for COBRA's dynamical model using experimental and simulated data, highlighting its potential to address sim-to-real gap in robot implementation.

Abstract

This paper employs a reinforcement learning-based model identification method aimed at enhancing the accuracy of the dynamics for our snake robot, called COBRA. Leveraging gradient information and iterative optimization, the proposed approach refines the parameters of COBRA's dynamical model such as coefficient of friction and actuator parameters using experimental and simulated data. Experimental validation on the hardware platform demonstrates the efficacy of the proposed approach, highlighting its potential to address sim-to-real gap in robot implementation.

Reinforcement Learning-Based Model Matching to Reduce the Sim-Real Gap in COBRA

TL;DR

A reinforcement learning-based model identification method aimed at enhancing the accuracy of the dynamics for COBRA's dynamical model using experimental and simulated data, highlighting its potential to address sim-to-real gap in robot implementation.

Abstract

This paper employs a reinforcement learning-based model identification method aimed at enhancing the accuracy of the dynamics for our snake robot, called COBRA. Leveraging gradient information and iterative optimization, the proposed approach refines the parameters of COBRA's dynamical model such as coefficient of friction and actuator parameters using experimental and simulated data. Experimental validation on the hardware platform demonstrates the efficacy of the proposed approach, highlighting its potential to address sim-to-real gap in robot implementation.
Paper Structure (13 sections, 15 equations, 12 figures)

This paper contains 13 sections, 15 equations, 12 figures.

Table of Contents

  1. Introduction
  2. Quick Overview of COBRA
  3. COBRA Full-dynamics
  4. Selection of Simulator
  5. Reinforcement Learning-guided Identification of Actuator and Stribeck Parameters
  6. Urgency
  7. Approach
  8. Performance Comparison Based on Underactuated and Actuted Dynamics from Eq. \ref{['eq:full-dynamics']}
  9. Reinforcement Learning Framework
  10. State, Action, and Reward Definition
  11. Policy Updates
  12. RESULTS
  13. Conclusion

Figures (12)

  • Figure 1: Illustrates COBRA platform while performing sidewinding over flat ground.
  • Figure 2: Shows free-body-diagram of COBRA and model parameters unknown to us and found as part of this work.
  • Figure 3: Highlights a significant disparity between the behaviors of COBRA in experiments and simulations. The paper's primary contribution lies in aligning the dynamic behavior of COBRA across both simulation and real-world experiments.
  • Figure 4: Reinforcement Learning-guided model identification setup used in this work.
  • Figure 5: Illustrates a comparison between the head positions in the actual hardware platform (blue), tuned model (orange) and untuned model (red) for a sidewinding trajectory 0.35, 0.5, and 0.65 Hz.
  • ...and 7 more figures