Currently in the lab

An office liaison that we built in public.

We bought a Unitree Go2 EDU+ in April 2026 and put it to work as the visitor concierge at our Castle Rock office. The Jetson Orin onboard runs a Whisper plus quantized 7B language model stack locally. Every win and every loss gets posted as a build log.

LIVE · Castle Rock office · last seen 14 minutes ago
FRAME 0024 · 30FPS CASTLE ROCK · CO SLAM LOCK · 0.34M UPTIME 04:18:22 QUADRANT III · INTENT: TRANSIT
Hero shot · placeholder · real photography to come
Telemetry · last 30 days
0 Days deployed
0 Conversations logged
0 Visitors escorted
0s Avg first response
0 Battery cycles
0 Emergency stops
Spec sheet

What is actually under the hood.

We list this because the buyers who matter ask. If your engineering team wants the full bill of materials, write us and we will send it.

Platform
  • Robot Unitree Go2 EDU+
  • Mass 15 kg
  • Battery 8000 mAh, 2 to 4 hr loaded
  • Compute NVIDIA Jetson Orin (onboard)
  • Sensors L1 3D LiDAR · depth · IMU · stereo
Software
  • Base Ubuntu 22.04 · ROS 2 Humble
  • Bridge micro-ROS over DDS
  • SDK Unitree SDK 2 (low + high level)
  • SLAM RTAB-Map with custom glass-mask layer
  • Container Docker, two services, one bridge
Agent layer
  • ASR Whisper-small.en · onboard
  • LLM Llama 3 8B Q4 · onboard fallback to Claude
  • TTS Piper · onboard
  • VLM Moondream 2 for visitor recognition
  • Orchestrator Custom Python · LangGraph patterns
Build logs

Weekly. Honest. Including the bits that did not work.

LOG 03

Why the L1 LiDAR hates a glass partition (and what we did about it)

The lobby has a frosted glass divider that the LiDAR reads as both a wall and a void depending on the angle. We disabled point returns under a confidence threshold and added a geofence the planner refuses to cross.

Read full entry
LOG 02

Wiring Whisper plus a local 7B model for sub-four-second turns

Cloud round trips averaged 2.1s before we even hit the LLM. We moved Whisper-small and a quantized Llama-3-8B onto the Orin. Mean turn time is now 3.8s, down from 6.4s. The trade-off is half the context window we used to have.

Read full entry
LOG 01

Getting micro-ROS talking to the Orin without the SDK fighting us

Unitree ships a Foxy-flavored stack. We wanted Humble for newer micro-ROS support. Two days of DDS config, one rebuilt kernel, one bridge container. The build instructions Unitree publishes are not wrong, just incomplete.

Read full entry
Want one of these in your space?

We can build you the same setup, branded for you.

If you saw what the office liaison does and thought "I want that for our lobby," we are the right team. Tell us about your space.