Undergraduate Researcher — Robotics group (reproducing ACT and Diffusion Policy)
Fine-tuningConducted reproduction work for imitation-learning papers (ACT and Diffusion Policy) with careful cross-checking of paper claims against repository/training behavior, using this to debug when training data or labels were the quiet failure source. Focused on identifying label/data quality issues such as noisy versus clean demonstrations and poor camera framing effects on downstream policy performance. Maintained a daily log of experiments and failure cases for advisor review to guide iterative improvements. • Compared repo behavior to paper claims to detect mismatches arising from dataset/label handling • Debugged training setup on a local RTX 4060 Ti 12GB under WSL2 to isolate whether labels or training inputs were problematic • Curated understanding of how demonstration cleanliness and camera framing influence model learning • Documented experiment attempts and breakages in a daily research log