Learning and Recognition of Force-based Robot Skills from Human Demonstrations

Contributor: Thomas Eiband

We develop algorithms that allow to recognize force-based robot skills learned by human demonstrations. Such skills require a contact with the environment and an in-contact control strategy, that can differ among several skills.

Haptic Exploration of Uncertain Workspaces

Humans naturally use their tactile abilitis to sense the environment and to efficiently manipulate objects. The same sensing modality can be employed by robots to extend their perception capabilites. Especially when relying on vision only, objects can be occlued or changes in lighting conditions can make a visual perception system unreliable. In our research, we proposed a technique how robots can make use of the sense of touch, called haptic exploration. It allows robots to detect changes in the environment and to adapt their actions accordingly, based on a tree of rigid and adaptive transformations.

Intuitive Programming of Conditional Tasks

Have you ever tried to manually program a system that shall make a decision? This can be a challenging task for non-experts that shall programm a robot without experience in the field. Therefore, we developed a framework that only needs to observe possible solutions to a decision making problem. We call such problems Conditional Tasks. The decision making is learned by the system and does not need to be implemented manually. This is achieved by monitoring the Conditional Tasks during execution and by switching to an alternative action, when a decision need to be made.

Understanding Force-based Robot Skills

Different actions are often transferred to a robot in the form of so-called skills. Such skills can be for example picking or placing an object. Whenever a skill is required to apply specific forces onto the environment, e.g. when grinding a surface with a tool, we consider them as force-based skills or contact skills. Hereby, our research goal is twofold:

  1. The robot can learn such type of skills from a human teacher
  2. The human teacher (or robot programmer) understands the capability of each of these skills and receives feedback what the robot has actually learned.

Related Publications

Eiband, T. & Lee, D.
Identification of Common Force-based Robot Skills from the Human and Robot Perspective
Humanoid Robots (Humanoids), IEEE-RAS 20th International Conference on, 2021

Qiu, Z.; Eiband, T.; Li, S. & Lee, D.
Hand Pose-based Task Learning from Visual Observations with Semantic Skill Extraction
2020 29th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), 2020, 596-603

Willibald, C.; Eiband, T. & Lee, D.
Collaborative Programming of Conditional Robot Tasks
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, accepted

Eiband, T.; Saveriano, M. & Lee, D.
Intuitive Programming of Conditional Tasks by Demonstration of Multiple Solutions
IEEE Robotics and Automation Letters, 2019, 4, 4483-4490

Eiband, T.; Saveriano, M. & Lee, D.
Learning haptic exploration schemes for adaptive task execution
IEEE International Conference on Robotics and Automation (ICRA), 2019, 7048-7054

Schoeffel, M.; Eiband, T. & Lee, D.
Exploiting internal and external force-sensing for compliance during force control
IEEE International Conference on Robotics and Automation (ICRA), Workshop on Learning for Industry 4.0: feasibility and challenges, 2019