Robot learning, computer vision and industrial AI — representations and rigorous evaluation.
M.Sc. in Computer Science, TU Darmstadt · Darmstadt, Germany.
I'm looking for research roles in industry and PhD positions.
More Structure, Not More Capacity: Object-Centric Representations for Visuomotor Imitation Learning
First author · RSS 2026 Workshop (non-archival) · OpenReview
| Project | Work |
|---|---|
| RARE26 | Early neoplasia detection in Barrett's esophagus: frozen GastroNet backbone, a 4096-parameter closed-form readout and a CPU-only inference container. Technical report · DOI |
| NeurIPS 2026 RoboSyn Challenge | Solo entrant; worked with ACT, vision-language-action models and world-action models. Three PRs merged into EmbodiChain. |
I built these tools while running research experiments with coding agents:
- wombat-gate: queues Git commits from sessions sharing a working tree, with each commit limited to named paths.
- dingo-pack: schedules heavy jobs on one machine with memory limits, and saves and restores coding sessions after a crash.