Done
AI for Industry Challenge 70 / 162
UR5e cable-port insertion project covering staged ROS 2 motion policy, Gazebo data collection, vision-based port localization, and retry logic.
View Details →
- Developed a ROS 2-based staged motion-planning policy node for UR5e cable-port peg-in-hole insertion.
- Tuned stiffness/damping parameters to stabilize contact response during insertion.
- Built a Gazebo automatic data-collection node and headless execution, initialization wait, ROS-Gazebo bridge, and stale-process cleanup pipeline to reduce repeated experiment bottlenecks.
- Integrated YOLO port recognition, multiview/stereo triangulation, and yaw/XYZ alignment logic to respond to port-position and approach-angle changes.
- Developed Vision and F/T sensor-based force baseline, offset prediction, and retry logic for insertion-failure recovery.
- Managed GitHub, Hugging Face Hub, and Notion to organize distributed experiment artifacts and collaboration flow.
- Handled teammate recruiting, schedule and role coordination, and experiment-flow documentation.
Done
Structural Stability Physical Reasoning AI Top 13.6%
Vision/Robotics competition project using image/video augmentation, multiview cross-attention, knowledge distillation, and ensemble calibration.
View Details →
- Implemented image/video augmentation and brightness/rotation preprocessing pipelines to increase dataset diversity.
- Implemented a multiview Bidirectional Cross-Attention model and preprocessing code to learn multiple structure viewpoints together.
- Compared key training conditions including backbone, epoch, and weight decay to secure stable physical-reasoning model performance.
- Configured learning stabilization with Knowledge Distillation, teacher regularization, mixup, EMA, and early stopping to transfer existing-model predictions and reduce overfitting.
- Recorded 66th place out of 484 teams (top 13.6%) through dev logloss analysis, temperature scaling, and ensemble weight search.