Learning-Based Monitoring of Non-Linear Mechanical Forces in Collaborative Robots
- Institute
- Lehrstuhl für Angewandte Mechanik
- Type
- Semester Thesis Master's Thesis
- Content
- experimental theoretical
- Description
1. Introduction
Collaborative robots (cobots) are increasingly used in industrial, service, and healthcare applications where they physically interact with humans and their environment. Unlike traditional industrial robots operating in structured settings, cobots must handle uncertain, non-linear mechanical processes such as:
-
Frictional effects in joints and contacts.
-
Flexibility and compliance in links, grippers, and end-effectors.
-
Backlash and hysteresis in transmissions.
-
Contact with deformable or variable materials.
These non-linearities make it difficult for conventional linear force models to reliably predict and monitor interaction forces. Failure to capture such effects can compromise safety, task accuracy, and robustness in collaborative operations.
Recent progress in machine learning (ML) and data-driven modeling opens opportunities to address these challenges by learning the non-linear force dynamics directly from sensory data.
2. Problem Statement
Traditional rigid-body dynamics with linear approximations are inadequate for modeling friction, flexibility, and compliance in cobot tasks. This results in:
-
Inaccurate force/torque predictions during contact-rich tasks.
-
Difficulty in compensating for non-linear mechanical behaviors.
-
Reduced reliability in monitoring for anomalies, such as unexpected collisions or excessive force during human-robot interaction.
Thus, there is a pressing need for a learning-based monitoring framework that captures and adapts to non-linear mechanical processes in real time.
3. Objectives
The proposed project aims to:
-
Model non-linear force dynamics caused by friction, flexibility, and compliance in cobot systems.
-
Develop learning-based methods (neural networks, Gaussian processes, recurrent models) to approximate these dynamics.
-
Implement a real-time monitoring system for detecting unsafe or abnormal force interactions.
-
Validate the framework experimentally on benchmark cobot tasks involving contact and compliance.
4. Methodology
4.1 Data Collection
-
Gather multi-modal data from cobot sensors (joint torque sensors, force/torque sensors, encoders). Starting with simulation.
-
Include scenarios with pronounced non-linearities:
-
Sliding and static friction in joints and contacts.
-
Flexible tool tips or soft grippers interacting with objects.
-
4.2 Learning-Based Modeling
-
Train models to learn the mapping between cobot states (position, velocity, torque) and measured forces.
-
Compare approaches:
-
Deep Neural Networks (DNNs): Non-linear function approximation.
-
Recurrent Neural Networks (RNNs) / LSTMs: Capturing time-dependent effects like friction hysteresis.
-
Gaussian Processes (GPs): Probabilistic modeling of uncertainty in force estimation.
-
4.3 Validation
-
Implement framework on a collaborative robotic platform (e.g., UR5, Franka Emika Panda).
-
Benchmark tasks:
-
Surface sliding with variable friction.
-
Human handover tasks with compliant forces.
-
Force-controlled assembly tasks.
-
5. Expected Outcomes
-
A validated learning-based force monitoring framework for non-linear mechanical processes in cobots.
-
Improved safety and robustness in collaborative operations.
-
Contribution to the development of adaptive, intelligent cobots capable of handling real-world uncertainties.
-
- Requirements
Background in machine learning and reinforcement learning. Basic understanding of physics behind Multi Link Manipulators. Experience with ML libraries (e.g., Python, TensorFlow, PyTorch). Basics in Simulink would be beneficial. Basics of English is required as thesis would be in English.
- Possible start
- sofort
- Contact
-
Tanmay Goyal, M.Sc
Room: 5501.03.104
Phone: +49 (89) 289 - 15226
tanmay.goyaltum.de