Undergraduate Research
Concordia College · Moorhead, MN
- Comparative study of imitation-learning policies (Action Chunking Transformer vs. SmolVLA) on a real dual-arm SO-101 robot, trained on identical demonstration data.
- Built the full pipeline end-to-end on Hugging Face's LeRobot framework — assembled and calibrated the hardware, collected a 200-episode (~71k-frame) teleop dataset, and published it to the Hugging Face Hub.
- Trained ACT from scratch (100k steps) and fine-tuned the 450M-parameter SmolVLA, then deployed both policies for autonomous execution on the physical arm.
- Diagnosed a deep stack of Windows GPU-training issues (PyTorch/torchcodec/FFmpeg alignment, native DLL load-order segfault, HF cache symlink permissions) to bring up training on a dual-RTX-4000-Ada workstation; profiled training to inform a single- vs. multi-GPU (DDP) scaling decision.