PhD Thesis: Combining interaction and perception to determine the physical properties of the robot environment
Teaching robots to "feel" their environment through deep learning - material classification, stiffness estimation, and terrain recognition using force/torque sensors.
Developed the transformer architecture for terrain classification on legged robots. Achieved 30x parameter reduction compared to CNN-RNN baselines while maintaining accuracy.
Designed a novel attention mechanism for fusing force/torque and IMU sensor data. Enables robust multi-modal perception for robotic manipulation.
Adapted Deep Embedding Clustering to the tactile sensing domain for exploratory material analysis without labeled data.
- Foundation Models for Tactile Sensing - Self-supervised learning for touch
- Terrain Classification - Cross-robot transfer, soft/deformable terrains
- Multi-Modal Fusion - Vision + touch integration