Visiting student researcher at Shanghai Jiao Tong University (advisor: Yutong Ban). Data infrastructure and training for vision-language-action models in bimanual manipulation.

Highlights

• Synchronized dual-arm joint states with 3-channel RGB at 30 FPS.
• 35% lower I/O latency across 100+ manipulation episodes.
• Scaled SFT and advantage-conditioned training for InternVL-1B with PyTorch FSDP on 8×A100 GPUs.
• +18% cross-task policy success.
• Human-in-the-loop teleoperation for closed-loop evaluation.

More context is on the Research page.