Stealth Startup

Robotics Software Engineer · Jan 2025 — Dec 2025

At a stealth robotics startup, I focused on vision–language–action (VLA) models and diffusion policies for dual xArm7 manipulation—closing the loop from data to reliable long-horizon behaviors on hardware.

I fine-tuned modern VLA families (including pi0 / pi0.5-class models) for our bimanual setup, and built synthetic data tooling that uses SAM2 and CUTIE-style mask tracking to augment real trajectories. I also improved diffusion policies by conditioning on segmentation targets, which lifted success on contact-rich pick-and-place.

On perception, I integrated FoundationPose and Grounded-SAM into a real-time stack aimed at sub-centimeter 6D grasping under clutter—bridging rough scene understanding with precise end-effector goals.

• VLA fine-tuning (pi0-class) for dual xArm7 long-horizon tasks
• Synthetic data pipelines with SAM2 / CUTIE-style mask tracking
• Diffusion policies with segmentation-aware targets for pick-and-place
• Real-time perception with FoundationPose + Grounded-SAM for precise 6D grasping