projects

Problem. Ricoh’s robotics group needed an object detector deployable on a Universal Robots 20 6-DOF industrial arm in a custom assembly cell. The objects of interest, the lighting, and the camera placement were all cell-specific, and real-world labeled data collection wouldn’t scale to the iteration cadence we needed.

Approach. Architected and built a four-stage synthetic-data pipeline end-to-end: OpenUSD scene authoring of the cell, Isaac Sim for physically-valid simulation, NVIDIA Cosmos as a diffusion foundation model for photorealistic seeding, and a modified YOLOv11 detector trained on the auto-labeled output. Bounding boxes are carried through from the scene graph rather than generated by a separate labeling pass, so every synthetic frame ships with ground-truth labels for free. Compute on Google Cloud H100s in Docker.

Outcome. Pipeline produced 10K+ auto-labeled synthetic frames in the operational run; detector deployed on the arm; iteration cost per new object class is orders of magnitude lower than a real-world capture campaign would be.

The longer architecture-level writeup is at /research/ricoh-synthetic-data.

Problem. Three threads at SEAKR (a satellite electronics shop, RTX subsidiary): future satellite hardware needed reliable boot software for critical memory systems; existing satellite mission software had kernel-level cybersecurity vulnerabilities that needed mitigation work; and a CNN-based machine-learning demo for software-defined radios needed to be brought up to a quality bar that could anchor a potential DARPA RFP.

Approach. Wrote low-level boot scripts for the memory subsystems, working against the constraints (radiation, certification, deterministic timing) that distinguish space-grade software from cloud-native code. Investigated the kernel-level vulnerabilities and designed mitigations against them. Debugged and upgraded the mission software applications across the satellite stack. On the ML side, refined the CNN demo on software-defined radios into a presentable artifact for the DARPA RFP context.

Outcome. Boot scripts integrated into the target hardware path. Vulnerability mitigations adopted into the mission software. SDR ML demo brought up to RFP-presentation quality and handed off to the team driving the proposal.

Problem. Lockheed Martin Space program executives needed a way to walk customers and stakeholders through full-scale 3D models of products without depending on physical prototypes or screen-bound mockups. Existing tooling didn’t preserve scale, didn’t carry component-level metadata at the right level of detail, and didn’t let multiple people inhabit the same scene at the same time.

Approach. Built an augmented-reality app from scratch that renders full-scale 3D models in physical space, with high-level descriptions of every component reachable by interaction. Then extended it with an infinitely expandable peer-to-peer network so multiple participants — anywhere — could share the same scene and walk through the model together in real time.

Outcome. Publicly demonstrated to a U.S. Air Force Major General, the Executive Vice President of Lockheed Martin Space, and multiple program executives. Final product presented to the Lockheed Martin board of directors, then handed off to a full-time team for continued development. Now used by the Pentagon to show military products to defense industry partners.