Anomaly Detection & RL for Robotic Assembly

M.Sc. research at the Technion: LSTM autoencoders for anomaly detection in real-world force/torque data and reinforcement learning for controller tuning, cutting assembly time by 37%.

Context: M.Sc. in Robotics, Technion – Israel Institute of Technology

Problem

Robotic assembly depends on contact-rich interaction, where small misalignments show up as subtle changes in the force/torque signal. Those failures had to be detected reliably from real sensor data, and the controller parameters that govern the assembly motion had to be tuned so the task ran faster without losing robustness.

Approach

Real-world data. I collected and preprocessed time-series data from a force/torque sensor on the robot's gripper to train the deep learning models.
Anomaly detection. I trained and deployed LSTM and LSTM autoencoder models to detect anomalies in the real-world time-series data.
Controller tuning with RL. I learned optimal controller parameters with reinforcement learning in a simulation environment.
Sim2Real. I transferred the learned parameters from simulation to the real robot and validated them on the physical assembly task.

Force/torque data feeds the LSTM anomaly detector, while RL-tuned controller parameters move from simulation to the real robot

Impact

37% reduction in assembly task duration on the real robot.
Anomaly detection running on real-world force/torque data from the robot.

Stack

Python · TensorFlow · PyTorch · MLflow · LSTM · Reinforcement Learning