[Paper Review] Learning Movement Assessment Primitives for Force Interaction Skills

Xiang Zhang, Athanasios Polydoros+1 more|arXiv (Cornell University)|May 11, 2018

Robot Manipulation and Learning11 references3 citations

TL;DR

This paper proposes Movement Assessment Primitives (MAPs), a task-agnostic, machine learning-based method for assessing robotic force-interaction skills via kinesthetic teaching. It uses Gaussian Process models to compare wrench-force profiles between demonstrations and reproductions, measuring similarity via Hellinger distance and classifying success probability with a Naive Bayes classifier—achieving up to 93.33% accuracy in industrial screwing and snap-fitting tasks.

ABSTRACT

We present a novel, reusable and task-agnostic primitive for assessing the outcome of a force-interaction robotic skill, useful e.g.\ for applications such as quality control in industrial manufacturing. The proposed method is easily programmed by kinesthetic teaching, and the desired adaptability and reusability are achieved by machine learning models. The primitive records sensory data during both demonstrations and reproductions of a movement. Recordings include the end-effector's Cartesian pose and exerted wrench at each time step. The collected data are then used to train Gaussian Processes which create models of the wrench as a function of the robot's pose. The similarity between the wrench models of the demonstration and the movement's reproduction is derived by measuring their Hellinger distance. This comparison creates features that are fed as inputs to a Naive Bayes classifier which estimates the movement's probability of success. The evaluation is performed on two diverse robotic assembly tasks -- snap-fitting and screwing -- with a total of 5 use cases, 11 demonstrations, and more than 200 movement executions. The performance metrics prove the proposed method's capability of generalization to different demonstrations and movements.

Motivation & Objective

• To develop a reusable, task-agnostic method for assessing robotic force-interaction skills without task-specific reprogramming.
• To enable non-expert users to teach robotic skills via kinesthetic demonstration and automatically assess their success.
• To generalize assessment across diverse demonstrations, initial states, and target positions in force-interaction tasks.
• To support integration into reinforcement learning by providing fast, data-efficient performance evaluation.
• To improve adaptability and reusability of robotic skills in industrial manufacturing and collaborative robotics.

Proposed method

• Collect end-effector pose and wrench data during human demonstrations and robot reproductions using a force-torque sensor.
• Train Gaussian Processes (GPs) to model the expected wrench as a function of robot pose, creating wrench models from demonstration and reproduction data.
• Compute the Hellinger distance between the GP-based wrench models of demonstration and reproduction as a similarity metric.
• Use the Hellinger distance as a feature input to a Naive Bayes classifier to estimate the probability of movement success.
• Train the MAPs using kinesthetic teaching and supervised learning, enabling reuse across different tasks without reprogramming.
• Apply leave-one-out cross-validation for evaluation across diverse movement states and demonstrations.

Experimental results

Research questions

• RQ1 Can a task-agnostic assessment primitive be learned from kinesthetic demonstrations to evaluate force-interaction skills without task-specific knowledge?
• RQ2 How well does the MAP method generalize across different demonstrations, initial states, and target positions in force-interaction tasks?
• RQ3 To what extent can the method detect failures in real-world robotic assembly tasks with minimal training data?
• RQ4 How effective is the Hellinger distance between GP-based wrench models in distinguishing successful from failed movements?
• RQ5 Can the MAP framework be reused across different robotic tasks (e.g., screwing vs. snap-fitting) without reprogramming?

Key findings

• The MAP method achieved 93.33% assessment accuracy on industrial snap-fitting tasks and 80% on industrial screwing tasks using leave-one-out cross-validation.
• Generalization performance was higher in industrial settings (93.33% for snap-fitting) than in toy-case experiments, due to fewer and more consistent failure modes.
• The method demonstrated data efficiency, correctly classifying failures even with only one training instance per failure class in the unbalanced snap-fitting dataset.
• Generalization to different demonstrations was more challenging, with performance dropping by about 10% compared to same-demonstration cases.
• The confusion matrices showed no class bias in screwing tasks, indicating robustness, while snap-fitting’s imbalance did not hinder accurate failure detection.
• The approach enables reusability across tasks without reprogramming, as only retraining is required when switching between tasks like screwing and snap-fitting.