Learning Pose Dynamical System for Contact Tasks under Human Interaction

Shangshang Yang, Xiao Gao, Zhao Feng, Xiaohui XiaoActuators 2023 · DOI

TL;DR

This paper introduces a new framework for robots to learn complex contact tasks from human demonstrations. It enables robots to not only replicate human movements and forces but also to safely interact with humans and seamlessly resume tasks afterward without needing to replan motion.

What Problem Does This Paper Address?

Robots are increasingly being used for repetitive contact tasks like polishing or laying up composite materials. A major challenge is enabling robots to learn these skills from human demonstrations, especially when precise control of both motion and force is required. Furthermore, when a human needs to interact with the robot during a task (e.g., to inspect or correct something), traditional methods struggle to allow the robot to resume its work smoothly without a complete re-planning of its motion. This paper tackles the problem of creating a unified framework that handles both learning complex contact tasks and robustly managing human interaction, allowing for seamless task resumption.

Methods

The core of the proposed framework is a time-invariant dynamical system (DS), specifically a pose diffeomorphic DS. This DS is used to model the human demonstration trajectory, allowing for online motion planning. The system learns both the position and orientation of the robot's end-effector. To handle the nuances of contact tasks and human interaction, the DS is augmented with admittance control. This allows the robot to modulate its motion and force output to meet different demands, such as reaching a contact space, performing the contact task, or complying with human interaction. The framework divides the task into three subtasks, each with a tailored motion controller.

Datasets

The paper does not explicitly mention using pre-existing public datasets for training. Instead, it describes a custom experimental setup for collecting demonstration data. This involves a UR5e robot, a force/torque sensor, and a platform designed for kinesthetic teaching. Two demonstrations were collected for each task: one for trajectory and one for force.

Key Insights and Ideas

  • Unified Framework for Learning and Interaction: The most significant insight is the development of a single framework that can learn complex contact tasks and gracefully handle human interaction, allowing for immediate task resumption.
  • Pose Diffeomorphic DS for Online Planning: Using a pose-aware, time-invariant DS with diffeomorphic mapping allows for efficient online motion planning, adapting to the learned human skills.
  • Admittance Control for Flexibility: The integration of admittance control provides the necessary flexibility to manage different phases of the contact task and to comply with human input during interaction.
  • Seamless Resumption After Interaction: A key advantage is the robot's ability to return to its demonstration trajectory immediately after human interaction, without requiring a full re-planning cycle.

Key Results

The experimental results with a UR5e robot performing a composite woven layup task demonstrate the effectiveness of the proposed framework.

  • The robot successfully reproduced the learned trajectory and force from human demonstrations.
  • The system allowed for safe human-robot interaction, with the robot complying with human input.
  • Crucially, after human interaction, the robot was able to automatically resume its task and return to the demonstration trajectory with small pose errors.
  • In experiments without interaction, the RMSE for position, orientation, and force were reported as 0.003 m, 1.78°, and 1.04 N, respectively, indicating good reproduction accuracy.
  • The method showed comparable force tracking performance to offline motion planning and DMP-based methods.

Limitations and Future Work

  • Computation Efficiency: The authors acknowledge that the online computation of mappings can be time-consuming. Future work will focus on developing strategies to improve computational efficiency.
  • Generalizability: While the framework is presented as generalizable to various contact tasks (polishing, wiping, pressing), further validation on a wider range of tasks and end-effectors would be beneficial.
  • Manual Parameter Tuning: Admittance controller parameters are obtained through manual adjustment. Future work could explore automated methods for parameter tuning.

Summary updated

Abstract

Robots are expected to execute various operation tasks like a human by learning human working skills, especially for complex contact tasks. Increasing demands for human–robot interaction during task execution makes robot motion planning and control a considerable challenge, not only to reproduce demonstration motion and force in the contact space but also to resume working after interacting with a human without re-planning motion. In this article, we propose a novel framework based on a time-invariant dynamical system (DS), taking into account both human skills transfer and human–robot interaction. In the proposed framework, the human demonstration trajectory was modeled by the pose diffeomorphic DS to achieve online motion planning. Furthermore, the motion of the DS was modified by admittance control to satisfy different demands. We evaluated the method with a UR5e robot in the contact task of the composite woven layup. The experimental results show that our approach can effectively reproduce the trajectory and force learned from human demonstration, allow human–robot interaction safely during the task, and control the robot to return to work automatically after human interaction.
BibTeX
@inproceedings{yang2023learning,
  author = {Shangshang Yang and Xiao Gao and Zhao Feng and Xiaohui Xiao},
  title = {{Learning Pose Dynamical System for Contact Tasks under Human Interaction}},
  booktitle = {Actuators},
  year = {2023},
  doi = {10.3390/act12040179},
}
Figure