OHMM-PA: A Learning from Demonstration Approach Using Online Hidden Markov Models with Path Planning
Jan Irsperger, Diego Fernandez Prado, and Eckehard Steinbach
AI summary
Problem
Learning from demonstration currently lacks stable solutions that offer high task generalization, industrial-level success rates, and minimal demonstration requirements while avoiding the need for continuous visual and haptic data or prior programming expertise.
Approach
The method decouples task execution into a coarse visual approach and a fine interaction phase, training an online Hidden Markov Model on kinesthetic demonstrations and using a custom path-planning algorithm to dynamically adjust trajectories based on real-time force and position feedback.
Key results
- 95% success rate across four industrial tasks with kinesthetic teaching
- Complex tasks learned in under 30 minutes with minimal demonstrations
- Online path algorithm dynamically replans trajectories to bypass obstacles or environmental changes
- Statistical fine-trajectory modeling compensates for inaccuracies in the coarse visual approach
Why it matters
Empowers non-expert industrial workers to rapidly and reliably program robots for contact-rich tasks without extensive programming knowledge or continuous sensor data.
Abstract
Learning from Demonstration (LfD) enables hu- mans to teach skills to robots quickly, without the need of knowledge in programming or robotics. Despite many proposed approaches, the field still lacks a stable solution with high task generalization, industrial-level success rates, and minimal demonstration requirements. In this work, a combination of Coarse-to-Fine (CtF) LfD with online Hidden Markov Models (HMMs) and impedance control is proposed to achieve higher task generalization with no prior task knowledge using only a handful of demonstrations. A coarse trajectory is generated to approach the task object using a self-supervised neural network with visual data. This is followed by an object interaction trajectory, employing the HMM approach with haptic data and a non-linear stiffness impedance controller. A new online HMM path algorithm is employed, demonstrating its ability to adapt to changes in the environment during execution. Experiments on a robotic arm show that complex tasks can be learned within 30 minutes and accomplish a 95 % success rate with kinesthetic teaching. The learned skill is stable and evaluated analytically.