Skip to content

architecture·

⏱ 90s

How neon fits inside the G1's runtime.

physical layout·

Unitree G1+
├─ Jetson Orin NX  (192.168.123.164)  ← runs neon
│    neon.service → .venv/bin/python agent.py   (or docker compose)
│      Strands loop · G1_ALL_TOOLS (53) · telegram/voice listeners
│    unitree_sdk2_python/  (local clone; pip wheel broken)
│    eth0 · CycloneDDS multicast
▼
└─ Main MCU        (192.168.123.161)  ← Unitree-owned, we never touch
     master_service → sport_mode · loco_service · arm_action · vui/audio
     │ SPI
     ▼  [29 motors · 2 arms · audio · LED · Livox lidar]

Key insight: neon and master_service don't know about each other. neon runs as a systemd unit (or docker compose) — if it crashes, motor control is unaffected; if sport_mode crashes, neon keeps running.

the agent loop·

flowchart LR
  M(["📨 message"]) --> I["inject live state<br/>(state·imu·battery·slam)"]
  I --> P["🎯 model plans tool calls"]
  P --> T["⚡ execute (parallel when independent)"]
  T --> R["💬 synthesize reply"]
  R --> U(["📤 user"])
  classDef a stroke:#cc5a3a,stroke-width:1.5px
  class I,P,T,R a

Every turn injects live robot state into the system prompt — the model wakes up already knowing FSM, IMU, battery, SLAM pose. All four reads hit in-memory DDS buffers (< 50 ms), not the motor bus.

four tool families·

flowchart LR
  A(["🤖 agent"]) --> C["🛡️ composed<br/>safety-gated"]
  A --> U["🔧 use_unitree<br/>universal SDK"]
  A --> D["📡 use_dds<br/>raw pub/sub"]
  A --> S["👁️ sensing<br/>camera·mic·slam"]
  C --> SDK["SDK clients"] --> DDS["CycloneDDS"]
  U --> SDK
  D --> DDS
  DDS --> MCU["⚙️ MCU"]
  S --> V["V4L2·ALSA"]
  classDef safe stroke:#6ee7b7,stroke-width:1.5px
  classDef raw stroke:#fbbf24,stroke-width:1.5px
  class C safe
  class D raw
  1. Composed — hand-written, safety-gated (FSM, mutex, clamps)
  2. use_unitree — AST-verified universal dispatch
  3. use_dds — raw escape hatch
  4. Sensing — outside the SDK (V4L2, whisper, YOLO)

one agent, four personas·

flowchart TB
  MEM[("shared memory<br/>kv · log · agent_log · briefings")]
  SH["🖥️ shell<br/>agent.py REPL"] --- MEM
  TG["💬 telegram<br/>listener"] --- MEM
  VC["🎙️ voice<br/>bidi + chest speaker"] --- MEM
  DP["🧵 dispatch<br/>cron sub-agents"] --- MEM
  classDef p stroke:#cc5a3a,stroke-width:1.5px
  class SH,TG,VC,DP p

All four share the same tools and the same memory. When one persona acts, the others see it via the unified reasoning log injected into their prompts — so voice can pick up where telegram left off. See voice architecture.

why DDS not ROS·

The MCU speaks CycloneDDS natively. ROS would add a bridge subprocess, an extra serialization hop, and middleware we don't control. Raw CycloneDDS + the SDK's typed wrappers is simpler, faster, and matches Unitree's own tools.

See tool catalog · systemd.