Low-Cost Robotic Manipulation

My Harvard ALM in Data Science capstone: an affordable robotic manipulation system built to make assistive pick-and-place accessible on a hobby budget.

Goal

Show that reliable, language-driven object manipulation does not require expensive industrial hardware. The full system runs on a ~$400 arm and an NVIDIA Jetson Nano.

How it works

• Perception uses a fine-tuned GQ-CNN grasp-quality model. Fine-tuning lifted grasp precision from 0.31 to 0.89 on the target objects.
• Reasoning uses a vision-language-action (VLA) policy so the system can take a natural-language instruction and act on it.
• Compute is fully on-device on the Jetson Nano, with the perception, planning, and actuation loop integrated and debugged end to end.

Result

96% task success across the evaluation set, with the whole pipeline running on low-cost, edge-deployable hardware.