Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comparison of Imitation Learning Algorithms for Bimanual Manipulation

Michael Drolet, Simon Stepputtis, Siva Kailas, Ajinkya Jain, Jan Peters, Stefan Schaal, Heni Ben Amor

TL;DR

This paper addresses the challenge of learning high-precision bimanual manipulation policies via imitation learning, evaluating several representative IL algorithms under noise and hyperparameter perturbations. The authors implement a two-armed UR5 bimanual insertion task in MuJoCo with an OSC-based expert and a 18D action/36D observation space, and compare BC, ACT, IBC, Diffusion Policy, GAIL, and DAgger. A three-phase experimental protocol—action/observation noise analysis, hyperparameter search with Optuna, and hyperparameter sensitivity evaluation using covering arrays—assesses data efficiency, robustness, and training demands. Key findings show that interaction-based methods and chunking-based methods (ACT and Diffusion) deliver robust performance under perturbations and require less training time, while GAIL tends to be data-hungry and sensitive to hyperparameters; BC and IBC lag in noisy settings. The work provides practical guidance for selecting IL methods for precision, contact-rich robotic manipulation and highlights the value of a careful experimental framework for benchmarking.

Abstract

Amidst the wide popularity of imitation learning algorithms in robotics, their properties regarding hyperparameter sensitivity, ease of training, data efficiency, and performance have not been well-studied in high-precision industry-inspired environments. In this work, we demonstrate the limitations and benefits of prominent imitation learning approaches and analyze their capabilities regarding these properties. We evaluate each algorithm on a complex bimanual manipulation task involving an over-constrained dynamics system in a setting involving multiple contacts between the manipulated object and the environment. While we find that imitation learning is well suited to solve such complex tasks, not all algorithms are equal in terms of handling environmental and hyperparameter perturbations, training requirements, performance, and ease of use. We investigate the empirical influence of these key characteristics by employing a carefully designed experimental procedure and learning environment. Paper website: https://bimanual-imitation.github.io/

A Comparison of Imitation Learning Algorithms for Bimanual Manipulation

TL;DR

This paper addresses the challenge of learning high-precision bimanual manipulation policies via imitation learning, evaluating several representative IL algorithms under noise and hyperparameter perturbations. The authors implement a two-armed UR5 bimanual insertion task in MuJoCo with an OSC-based expert and a 18D action/36D observation space, and compare BC, ACT, IBC, Diffusion Policy, GAIL, and DAgger. A three-phase experimental protocol—action/observation noise analysis, hyperparameter search with Optuna, and hyperparameter sensitivity evaluation using covering arrays—assesses data efficiency, robustness, and training demands. Key findings show that interaction-based methods and chunking-based methods (ACT and Diffusion) deliver robust performance under perturbations and require less training time, while GAIL tends to be data-hungry and sensitive to hyperparameters; BC and IBC lag in noisy settings. The work provides practical guidance for selecting IL methods for precision, contact-rich robotic manipulation and highlights the value of a careful experimental framework for benchmarking.

Abstract

Amidst the wide popularity of imitation learning algorithms in robotics, their properties regarding hyperparameter sensitivity, ease of training, data efficiency, and performance have not been well-studied in high-precision industry-inspired environments. In this work, we demonstrate the limitations and benefits of prominent imitation learning approaches and analyze their capabilities regarding these properties. We evaluate each algorithm on a complex bimanual manipulation task involving an over-constrained dynamics system in a setting involving multiple contacts between the manipulated object and the environment. While we find that imitation learning is well suited to solve such complex tasks, not all algorithms are equal in terms of handling environmental and hyperparameter perturbations, training requirements, performance, and ease of use. We investigate the empirical influence of these key characteristics by employing a carefully designed experimental procedure and learning environment. Paper website: https://bimanual-imitation.github.io/
Paper Structure (19 sections, 6 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 19 sections, 6 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Algorithm Selection
  4. Behaviorial Cloning
  5. Action Chunking Transformer
  6. Implicit Behavorial Cloning
  7. Diffusion Policy
  8. Generative Adversarial Imitation Learning
  9. DAgger: Dataset Aggregation
  10. Methodology
  11. Operational Space Controller (OSC)
  12. Bimanual Manipulation Expert Controller
  13. Environments
  14. Experimental Setup
  15. Action and Observation Noise Analysis
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Two UR5 arms equipped with grippers and mounted to a rotating torso. The robot above completes the final stage of the high-precision four-peg insertion task.
  • Figure 2: The robot successfully transfers and inserts the dynamic adapter (white) into the stationary adapter (black).
  • Figure 3: The Wasserstein distance, as measured by the state-action samples from the best policies in the Zero Noise environment with 200 expert demonstrations.
  • Figure 4: A high-level interpretation of the key metrics describing the algorithms in the bimanual insertion environments.
  • Figure : Bimanual Insertion Expert