Research Intern

• Lead task design for a long-horizon human–robot collaboration benchmark, developing simulated cooking scenarios, physical success checks, and partial-information variants; built an early state-engine prototype and found that symbolic task constraints masked visual activity-recognition errors.

Research Intern · Advisor: Prof. Ziwei Liu

• Co-led EgoTools, a dataset and benchmark for tool-centric reasoning: collected ~100 hours of egocentric tool-use video and designed a 1K-QA benchmark across four reasoning tracks, with annotation and quality-control workflows.
• Ran input ablations validating the benchmark’s reliance on visual and temporal evidence; contributed to the Qwen3-VL-8B fine-tuning study (EgoTools-Bench 50.0% → 60.9%) and its cross-benchmark generalization analysis.

Research Intern · Advisor: Prof. Seungryong Kim

• Jun 2026 – Oct 2026: For DreamX, ran the latent-action baseline comparison on LARYBench under a shared protocol, and co-ran human-to-robot action-transfer experiments on a bimanual robot, recording human demonstrations and paired teleoperated reference trajectories.
• Jun 2025 – Nov 2025: Contributed to C3G, analyzing DINOv2-based attention maps to study how compact Gaussian queries aggregate multi-view information.

Research Intern · Advisor: Prof. Mun Churl Kim

• Implemented and evaluated object-detection pipelines spanning two-stage, YOLO-style, and Vision Transformer architectures for few-shot and anomaly detection tasks.