We propose an adaptive walking framework that combines a friction-aware Model Predictive Control (MPC) strategy with a vision-based friction estimation module. The controller is based on the Linear Inverted Pendulum Model (LIPM) and explicitly considers slippage risks by incorporating friction constraints. A learned vision model estimates ground friction from camera input, enabling the robot to adapt its gait in real time to changing terrain conditions. The system will be validated in simulation and potentially on hardware, integrating control, perception, and learning.

• Background: Classical walking pattern generators assume a fixed, often overly conservative, friction coefficient for foot-ground interaction, which limits their performance on varying terrains. The paper “Pattern Generation for Walking on Slippery Terrains” introduces a friction-aware LIPM-MPC, allowing the planner to explicitly reason about slippage risk. Separately, learning-based approaches such as “Identifying Terrain Physical Parameters from Vision” demonstrate how terrain properties like friction can be estimated from visual cues using learned models.

• Objective: This thesis aims to integrate a learned vision-based friction estimator into the walking controller. This allows the robot to adjust its walking strategy according to terrain appearance, supporting safer and more agile locomotion in real-world environments.

• Enrolled in a Computer Science, Robotics, or related MSc program (Bonn/Cologne area)
• Strong background in linear algebra, optimization, and/or machine learning
• Experience with Python, and optionally C++, PyTorch/TensorFlow
• Motivation for real-world robot deployment and interest in publishing scientific results

1. Khadiv, M. et al., “Pattern Generation for Walking on Slippery Terrains,” ICRoM, 2017.
2. Chen, J. et al., “Identifying Terrain Physical Parameters from Vision—Towards Physical-Parameter-Aware Locomotion and Navigation,” IEEE RA-L, 2024.