Projects

Role: Owned the pickup-and-putdown loop around SO-ARM101 data collection, LeRobot-style dataset organization, model training/deployment validation, and failure analysis.

Pipeline: teleoperation -> LeRobotDataset -> ACT/Diffusion Policy/SmolVLA training -> OpenPI/pi0/pi0.5 inference integration -> real-robot rollout.

Evidence: public repo xiaoms22/soarm101, 99 episodes / ~28k frames, model rollout notes, and pi0/pi0.5 small-sample deployment validation.

Boundary: pi0/pi0.5 work is deployment validation and small-sample evaluation, not from-scratch foundation-model training.

Next: Expand object, position, lighting, camera-view, and language-paraphrase tests into a more standard generalization table.

Role: Built parts of a teaching-training-deployment prototype for a dual-arm hazardous-material placement task.

Pipeline: ROS2 nodes/UI -> visual and joint-state collection -> state-action schema -> LeRobot interface -> demonstrations -> model training validation.

Evidence: public summary only; no internal files, raw project code, or partner materials are uploaded.

Boundary: Prototype pipeline and training validation, not final industrial delivery.

Next: Improve dual-arm synchronization, motion safety constraints, failure taxonomy, and evaluation records.

Role: Supported THMOS in RoboCup China humanoid Small group and won silver / second place.

Pipeline: High Torque XiaoPai humanoid robots -> network/IP and remote-controller binding -> startup flow -> finite-state-machine debugging -> field troubleshooting.

Evidence: competition result and real robot deployment/debugging experience.

Boundary: Field deployment and system debugging contribution, not full vision or gait algorithm ownership.

Next: Convert field issues into startup checklists, network setup notes, and state-machine troubleshooting records.

Role: Worked on the perception-decision-action loop for a small embedded robot.

Pipeline: sensor input -> FSM task logic -> PID tuning -> obstacle avoidance, path planning, pushing, and shooting.

Evidence: supporting evidence for low-level control, embedded debugging, and task-state design.

Boundary: Robotics/control engineering project, not an embodied foundation-model project.

Role: Built experience across scientific simulation, physical-system modeling, multimodal spatial registration, and ROS2-based robot software integration.

Pipeline: theoretical modeling -> numerical simulation -> parameter scan -> error analysis -> engineering-constraint reasoning.

Evidence: supporting evidence for mathematical modeling, scientific computing, and complex system understanding.

Boundary: Background evidence for modeling and systems skills; the main embodied-AI evidence is SO-ARM101 and the dual-arm project.