Research
Marso Hack Berlin 2026 – Robot Parcel Sorting Challenge
07/2026Imitation learning for simulated parcel sorting with ManiSkill3, SAPIEN, and a Franka Panda manipulator.
- Completed an imitation-learning solution for a simulated robot parcel-sorting challenge using ManiSkill3, SAPIEN, and a Franka Panda manipulator.
- Parsed and replayed expert demonstration trajectories stored in HDF5/JSON formats across easy, medium, and hard task settings, extracting observations, actions, environment states, and success labels for training and evaluation.
- Established behavior-cloning baselines from state and RGB observations and evaluated State Diffusion Policy and RGB ACT policies in local rollout episodes.
- Built the training and rollout pipeline and measured local policy performance by task success rate.
SO-101 Real-World Robotic Learning with LeRobot
07/2026End-to-end imitation learning on the Huaner SO-101 dual-arm platform using LeRobot.
- Reproduced an end-to-end imitation-learning pipeline on the Huaner SO-101 dual-arm platform using the LeRobot framework, covering teleoperation data collection, SmolVLA policy training, and real-world deployment.
- Configured the full software stack and performed source-level modifications to enable RGB multi-view perception and language-conditioned manipulation.
- Diagnosed and resolved hardware setup, camera calibration, and inference failure cases through iterative debugging and validation.
YOLOv8 and IoT Agricultural Pest Detection System
01/2026Undergraduate thesis combining embedded sensing, live video, and YOLOv8 pest detection.
- Designed the embedded control architecture on an STM32F103C8T6 microcontroller, integrating DHT22 temperature and humidity sensing with an effective accumulated-temperature algorithm for automated SMS pest-outbreak alerts.
- Configured an ESP32-S3-WROOM-1U WiFi module to stream live field video from the microcontroller node to a host-computer gateway for real-time processing.
- Built the video ingestion and preprocessing pipeline with OpenCV to feed live field footage into the inference pipeline.
- Fine-tuned and deployed a YOLOv8x model for real-time pest detection on the IP102 dataset, with results visualized through a PyQt5 interface backed by SQLite.
MSM-Seg: Multi-Modal Brain Tumor Segmentation
03/2025Comparative 3D segmentation experiments on BraTS 2024 and BraTS 2025 for the MSM-Seg framework.
- Conducted comprehensive comparative experiments by training and evaluating 3D segmentation models (nnU-Net, 3D-TransUNet, TransBTS, SIN2GR, UNETR++, and others) on BraTS 2024 and BraTS 2025 datasets.
- Implemented standardized evaluation protocols in the BraTS-METS benchmark, measuring Dice and HD95 for ET, TC, and WT sub-regions.
- Analyzed quantitative results to validate that MSM-Seg outperforms or matches all baselines, particularly under the category-agnostic prompting setting.