I'm working on robot learning and deployment engineering — training policies that give robots new skills, and building the infrastructure that takes them from lab demo to the real world. Feel free to reach out if you'd like to chat!
Currently
- Fine-tuning an open-weights VLA on data from our own cell and evaluating it against a scripted baseline
- Temporal-grounding evals for VLAs — found duration-blindness across π0.5, π0-FAST, GR00T N1.x, MemoryVLA (~255 episodes); small conditioning fix worth +6pp zero-shot. Writeup in progress.
Some things I've worked on
- Bimanual UR5e work cell, end to end — Go control plane owning robot drivers, camera health, and cell telemetry
- Calibration stack — hand-eye extrinsics, per-dock re-registration, calibration-drift monitoring
- Teleop data engine — GELLO-style rig, MCAP recording, automated quality gates, LeRobot-format export
- Multi-GPU VLA fine-tuning on cell-collected data (π0.5), including spot-interruption handling and eval harnesses
- HIL-SERL-style intervention gating — space-mouse takeover logic during policy rollouts
- Integrated a vendor WebRTC teleop stack (Sentinel) and wrote the 125 Hz chunk-consumer interpolation layer on top
- Distillery — egocentric-video annotation pipeline at cloud scale
- Javelin — pgvector semantic search, Go+SQS data pipelines, structured-output SSE streaming