Jelle Dehn

Robot-arm learning and control

I'm working towards precise movement with an SO-101 robot arm, using reinforcement learning in MuJoCo and transfer to the physical robot. The controller learns motor duty-cycle commands directly. Compared with the excavator in my thesis, the arm has less actuation delay. Current work focuses on movement accuracy and investigating intermittent sensor-read failures on the hardware.

I adapted SmolVLA a vision-language-action (VLA) model to output Cartesian actions and use custom observation selection, connecting its outputs to my low-level control policy.

GitHub: robot_arm

Excavator control

My MSc thesis investigated excavator control with a target precision of 3 cm over a reach of roughly 6 metres. Hydraulic delays mean commands can take effect several control steps later, making the current observation alone insufficient for control. I compared sequence models and fixed history windows to learn a state representation for Soft Actor-Critic (SAC), evaluating auxiliary tasks including state and action prediction.

I ported the controller to MATLAB/Simulink, and Bosch deployed a learned policy on a physical 12-tonne excavator.

LLM file-editing agent

Built an LLM coding assistant with LangGraph, implementing tools for file access and code editing.

GitHub: llm-gym