← Back ICRA 2026

Task-Parameterized Motion Learning with Time-Sensitive Constraints

Julian Richter, João P. Oliveira, Christian Scheurer, Jochen J. Steil, Niels Dehio

PDF

AI summary

Key figure (auto-extracted from paper)

Enforcing time-sensitive constraints directly during Gaussian Mixture Model learning guarantees precise start, via, and end pose execution in robot motion skills.

Motion learning Gaussian mixture models Constrained optimization Task-parameterization Robotics Learning from demonstration

Problem

Standard learning-from-demonstration methods using Gaussian Mixture Models fail to guarantee precise execution of critical poses at specific times, limiting their transfer to industrial applications. Existing constraint-enforcement techniques typically apply only at inference time, compromising learning efficiency and accuracy.

Approach

The authors introduce a Constrained Expectation Maximization (CEM) algorithm that modifies the GMM learning process to enforce time-sensitive constraints directly during training. The method operates on Riemannian manifolds and extends to task-parameterized scenarios to handle variable motion skills.

Key results

Novel CEM algorithm for GMMs on Riemannian manifolds
Extension of CEM to Task-Parameterized GMMs (TP-GMM)
Improved and more efficient reproduction of demonstration data compared to inference-time constraint methods
Successful validation on handwritten data and real KUKA LBR iiwa pick-and-place tasks

Why it matters

Enables reliable, precise motion transfer for non-expert robot programming by guaranteeing exact pose execution at critical timestamps, directly addressing a key barrier to industrial adoption of learning-from-demonstration.

Abstract

Teaching motion skills to robots through demonstra- tions has becomes widely popular. However, precise execution of start-, via-, and end-poses at given times is often not guaranteed, limiting the technology transfer to industrial application. To address this issue, we propose the novel Constrained Expectation Maximization (CEM) algorithm, which enforces time-sensitive constraints (TSC) when learning Gaussian Mixture Models (GMM). Our approach applies to data on Riemannian manifolds and extends to task-parameterized scenarios. We validate CEM against state-of-the-art methods on handwritten data and real robot applications utilizing the KUKA LBR iiwa. By enforcing constraints within the learning process, CEM achieves improved and more efficient reproduction of the demonstration data.

Index terms

Learning from Demonstration Constrained Motion Planning