Detecting Obstacles by Predicting Motor Effort

In this seminar, I will present my ongoing work on how a robot can infer obstacles by predicting the effort required for movement. Epi, currently moves its motors based only on target positions, without adjusting for physical resistance or controlling force. To address this, I am developing models that predict how much electric current the motors should need to reach different positions. If the actual current is unexpectedly high, this could indicate that the robot has encountered an external obstacle.

This approach could eventually be integrated with visual input to improve performance, similar how humans combine sensory modalities to understand their own movements. I started with simple Bayesian regression models to map motor positions to expected current and am now training neural networks for more detailed and adaptive predictions. However, my models tend to underestimate the required current, and I hope to use this seminar to gather feedback and ideas for improvement.

___
Open for external guests.

Please request access to Zoom link in advance if participating virtually.