A 12-DOF quadruped ("robot dog") you can drive from a CLI / HTTP endpoints over WiFi or BLE, with a locomotion policy trained in simulation, plus a camera-trigger that signals an external service to take a photo on the robot's behalf.

The original robot/RobotDogUpdt.f3d is an Autodesk Fusion 360 archive (proprietary binary B-Rep; Fusion does not run on Linux). The robot model here is a fresh parametric URDF/MJCF re-derived from scratch — see model/.

8-DOF build: an ESP32 drives 8 servos through a PCA9685 over I²C. Servos are powered from a dedicated BEC (never from the ESP32). Right-side legs (FR, RR) are mirror-mounted — handled in servo_map.json (direction = -1), not in wiring.

flowchart LR
  BAT["Battery + switch + fuse"] --> BEC["BEC/UBEC 5-6V, min 5A"]
  BEC --- CAP["470-1000uF cap across V+/GND"]
  ESP["ESP32 dev board<br/>powered via USB or 5V/VIN"]
  PCA["PCA9685 16-ch PWM<br/>I2C addr 0x40, 50 Hz"]
  ESP -- "GPIO21 SDA" --> PCA
  ESP -- "GPIO22 SCL" --> PCA
  ESP -- "3V3 to VCC logic" --> PCA
  BEC -- "V+ servo rail" --> PCA
  PCA -- ch0 --> A0["FL_abd"]
  PCA -- ch1 --> A1["FL_knee"]
  PCA -- ch2 --> A2["FR_abd (mirror)"]
  PCA -- ch3 --> A3["FR_knee (mirror)"]
  PCA -- ch4 --> A4["RL_abd"]
  PCA -- ch5 --> A5["RL_knee"]
  PCA -- ch6 --> A6["RR_abd (mirror)"]
  PCA -- ch7 --> A7["RR_knee (mirror)"]
  ESP -. "common GND" .-> GND["Star ground:<br/>battery = BEC = PCA9685 = ESP32"]
  BEC -. "common GND" .-> GND
  PCA -. "common GND" .-> GND

The policy/gait runs on a companion computer (DGX Spark), not the ESP32. Commands go over WiFi to the ESP32; for scan runs the DGX also pulls photos from a phone (IP Webcam) into gaussian_splatting_photos/run_N/.

flowchart TB
  subgraph DGX["DGX Spark (GB10) - companion computer"]
    TRAIN["Isaac Lab training"] -->|exports| ONNX["policy_isaac.onnx"]
    ONNX --> PR["policy_runner.py<br/>ONNX inference / gait"]
    RC["robotctl.py (CLI)"]
    CC["capture_coordinator.py<br/>dead-reckon ~1cm trigger"]
    CC -->|writes| GS[("gaussian_splatting_photos/<br/>run_N : jpg + manifest")]
  end
  ESP["ESP32 firmware<br/>HTTP /cmd, WS /ws<br/>mDNS robotdog.local"]
  PCA2["PCA9685"] --> SERVOS["8 servos<br/>4 legs x abd+knee"]
  PHONE["Phone on robot<br/>IP Webcam :8080<br/>+ WiFi hotspot AP"]
  PR -->|"set_joints/gait over WiFi"| ESP
  RC -->|"commands over WiFi"| ESP
  ESP -->|"telemetry /state"| PR
  ESP --> PCA2
  CC -->|"GET /shot.jpg over WiFi"| PHONE

python3 -m venv .venv && . .venv/bin/activate
pip install mujoco numpy pyyaml requests websockets aiohttp bleak \
            stable-baselines3 onnx onnxruntime onnxscript pillow

python model/generate.py            # build URDF/MJCF/servo_map/servo_config.h
python sim/test_servo_map.py        # validate servo calibration round-trips
python sim/view_mujoco.py --check   # validate kinematics
python sim/walk_demo.py             # open-loop trot — robot walks forward

# drive a simulated robot end-to-end:
python control/mock_esp32.py 8770 &                       # fake ESP32
python control/cli/robotctl.py --host 127.0.0.1 --port 8770 stand
python control/cli/robotctl.py --host 127.0.0.1 --port 8770 walk --speed 0.6

1. Wire per firmware/README_WIRING.md (ESP32 + PCA9685 + 12 servos + BEC).
2. Set WiFi creds in firmware/src/config.h, then cd firmware && pio run -t upload.
3. Control: robotctl --host <robot-ip> stand (HTTP), --ws (WebSocket), or --ble. The robot also raises its own AP robotdog/walkies123 if it can't join WiFi.

• CLI: control/cli/robotctl.py — stand, stop, estop, walk, gait, set-joint, joints, photo, state. Transports: HTTP (default), --ws, --ble.
• HTTP endpoints (served by the ESP32): POST /cmd (JSON envelope), GET /state.
• WebSocket /ws — for 50 Hz joint streaming (the policy runner).
• See docs/command_protocol.md for the exact schema.

• Primary (GPU): Isaac Lab task on the NVIDIA GB10 — training/isaaclab_task/, driven by training/remote/{setup_gb10,sync,train}.sh. See docs/isaac_lab_gb10.md.
• Fallback (CPU/anywhere): training/mujoco_fallback/train_ppo.py — same obs/action layout, so policies are interchangeable.
• Deploy: training/export_policy.py → policy/policy.onnx, then control/runtime/policy_runner.py --sim (closed-loop in MuJoCo) or --robot <ip> (stream to the ESP32). The policy runs on a companion computer, not the ESP32 — see docs/sim2real.md.

The robot does not own a camera. control/runtime/camera_trigger.py (host) or the firmware's photo command fires a simple HTTP request to an external capture service (the device holding the camera / DJI drone / phone). Set the webhook URL in camera_trigger.py or pass --webhook to robotctl photo.