Chair of Sensor Based Robotic Systems and Intelligent Assistance Systems

The primary mission of this chair is teaching and research in the fields of design, sensorbased programming, and control of complex robotic systems for manipulation and locomotion. The main field of application is the robotic assistance in industrial production, medical and health-care applications up to personal assistance systems.

The chair is in close connection with the Institute of Robotics and Mechatronics of the German Aerospace Center (DLR).

For open docorate positions please contact us.

Announcement of Scientific Presentation

Title: Interactive Robot Skill Learning and Adaptation Through Probabilistic Movement Primitives and Natural Language Integration

Presenter: Markus Knauer

Abstract

This work addresses the challenge of data-efficient and intuitive robot skill acquisition and adaptation through a unified framework combining probabilistic movement representations, interactive imitation learning, and natural language understanding.

 We present three complementary approaches to robot skill learning: (1) an interactive incremental learning framework leveraging Task-Parameterized Kernelized Movement Primitives  (TP-KMPs) that enables local trajectory modulation through direct physical human feedback, allowing robots to improve generalization, interactively add new task parameters, and extend skills beyond demonstrated regions with uncertainty-aware stiffness regulation; (2) a tool-based architecture (IROSA) that enables zero-shot natural language-driven skill adaptation by  maintaining a protective abstraction layer between pre-trained language models and robot hardware, allowing users to modify execution speed, correct trajectories, and handle obstacle avoidance without model fine-tuning; and (3) a comprehensive methodology integrating active learning formulations, uncertainty quantification distinguishing between aleatoric and epistemic uncertainty, and LLM-based skill adaptation mechanisms for autonomous skill selection and parameter tuning.

 These approaches share a common foundation in probabilistic, non-parametric skill representations that encode both data variance and model uncertainty, enabling flexible trajectory modulation while maintaining interpretability, safety, and predictable robot behavior. Experimental validation on a 7-DoF torque-controlled robot performing industrial bearing ring insertion tasks demonstrates reliable skill adaptation through physical corrections, interactive object addition, and natural language commands, achieving superior generalization compared to baseline methods while enabling immediate real-world deployment in industrial settings.

_______________________________________________________________

Zoom
24. November 2025 2:00 PM
https://tum-conf.zoom-x.de/j/69076268562?pwd=sHWF8ziXm595P2WSB7Pv0dD2dbktBc.1
Meeting-ID:  690 7626 8562
Kenncode: 133277