Turn a 3D skeleton into a pseudo-image, then retrieve it like a document — with token-level explainability.
News
- Mar 2026 Three new preprints on motion retrieval, motion representation, and e-textile design are now available on arXiv.
- Nov 2024 Started my doctoral research at Aalto University, funded by AI-DOC.
Motion Understanding, HAR & E-Textile
What if human motion had its own alphabet, words, and grammar — just like natural language?
LingoMotion: An Interpretable and Unambiguous Symbolic Representation for Human Motion
Enabling models to continuously learn new activities from streaming sensor data without forgetting.
Efficient Online Continual Learning in Sensor-Based Human Activity Recognition
Can an LLM tell you where to sew a sensor onto a shirt? Turns out, it depends on who's asking.
Exploring Human-AI Collaboration in E-Textile Design: A Case Study on Flex Sensor Placement for Shoulder Motion Detection
Which knitting pattern makes the best wearable strain sensor? Plated knitting wins — with gauge factors above 13.
Comparative Analysis of Knitting Patterns for Wearable Resistive Strain Sensors in Joint Motion Measurements
Vision-Language Retrieval
What if your query encoder didn't need to see at all? 2B → 70M, 95% quality, 50× faster.
NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval
Everyone looks at the last layer. We looked in the middle — and pruned 90% of index vectors without losing retrieval quality.