mechdog

name	mechdog
description	Control a Hiwonder MechDog robot or the local MechDog simulator through tool calls.
tags	["robotics","simulator","mechdog"]
metadata	{"openclaw":{"emoji":"robot","requires":["MECHDOG_IP for hardware, optional for simulator","OPENAI_API_KEY for OpenAI vision"]}}

MechDog Control Skill

Use this skill to control a Hiwonder MechDog quadruped or the local simulator.

Operating Rules

• For movement or action requests, call a MechDog tool directly.
• For natural-language movement requests, first derive the exact low-level command list that will execute the request, then run that list.
• If direct MechDog tools are not exposed in the current runtime, use the exec tool to call the simulator or robot HTTP API directly.
• In exec fallback mode, only use the real transport endpoints: /move, /action, /status, /position, /reset, and camera endpoints. Do not invent HTTP endpoints from tool names such as /circle, /square, or /movement_pattern.
• Do not use sessions_spawn for MechDog movement, action, status, camera, or vision requests. The MechDog agent should execute the skill itself instead of asking a subagent to control the same dog.
• Do not emit multiple MechDog tool calls in the same assistant response when order matters. Use one routine tool when available, or execute steps strictly one at a time.
• Do not manually improvise circle, square, triangle, figure-eight, zigzag, patrol, or loop requests. Convert them into an explicit command list first, then execute the list.
• For rich natural-language requests that combine patterns, timing, and follow-up actions, prefer execute_request so the LLM can plan once and the skill can execute deterministically.
• Do not ask the user for simulator credentials or access details for normal local simulator use.
• If MECHDOG_IP is unset, assume the simulator at localhost:3000.
• Keep tool usage concrete and brief. Prefer one tool call per requested step.

Tool Routing

move

Use move for locomotion and turning commands.

Parameters:

• direction: one of forward, backward, left, right, stop
• duration_ms: optional, default 1000

Interpretation:

• "move left", "go left", "turn left", "rotate left" -> direction: "left"
• "move right", "go right", "turn right", "rotate right" -> direction: "right"
• "move forward", "walk forward", "go ahead" -> direction: "forward"
• "move backward", "go back", "reverse" -> direction: "backward"
• "stop", "halt" -> direction: "stop"

Defaults:

• If the user gives no duration, use 1000 ms.
• If the user says "for 3 seconds", use 3000 ms.

Examples:

• "Ask the dog to move to the left." -> call move with direction: "left", duration_ms: 1000
• "Walk forward for 2 seconds." -> call move with direction: "forward", duration_ms: 2000

HTTP fallback:

curl -sS -X POST "http://${MECHDOG_IP:-localhost:3000}/move" \
  -H "Content-Type: application/json" \
  -d '{"direction":"forward","duration":1000}'

action

Use action for preset actions.

Parameters:

• action: one of sit, stand, shake, wave, dance, balance

Examples:

• "Sit down." -> action: "sit"
• "Stand up and wave." -> first action: "stand", then action: "wave"

HTTP fallback:

curl -sS -X POST "http://${MECHDOG_IP:-localhost:3000}/action" \
  -H "Content-Type: application/json" \
  -d '{"name":"sit"}'

look

Use look to capture a camera frame without analysis.

square_pattern

Use square_pattern for square, box, perimeter, loop, patrol, or multi-side walking requests.

Parameters:

• segment_duration_ms: walk time for each side
• turn_direction: left or right
• final_actions: optional follow-up actions such as sit and wave

Examples:

• "Move in a square pattern for 5 seconds each direction and then sit and wave." -> call square_pattern with segment_duration_ms: 5000, turn_direction: "right", final_actions: ["sit", "wave"]
• "Walk a box pattern, 2 seconds per side, turning left." -> call square_pattern with segment_duration_ms: 2000, turn_direction: "left"

square

Use square for direct requests that literally ask to move in a square or box.

Examples:

• "Move in a square for 5 seconds each side." -> call square with duration_ms: 5000
• "Do a left square and then sit." -> call square with turn_direction: "left", final_actions: ["sit"]

circle_pattern

Use circle_pattern for circle, lap, orbit, round loop, or curved patrol requests.

Parameters:

• segment_duration_ms: total time for one full circle
• turn_direction: left or right
• repeat_count: optional number of full circles
• final_actions: optional follow-up actions such as sit and wave

Examples:

• "Move in a circle for 10 seconds." -> call circle_pattern with segment_duration_ms: 10000, turn_direction: "right"
• "Do two left circles and then wave." -> call circle_pattern with segment_duration_ms: 4000, turn_direction: "left", repeat_count: 2, final_actions: ["wave"]

circle

Use circle for direct requests that literally ask to move in a circle, lap, orbit, or loop.

Examples:

• "Move in a circle for 10 seconds." -> call circle with duration_ms: 10000
• "Do two circles and then wave." -> call circle with repeat_count: 2, final_actions: ["wave"]

command_plan

Use command_plan when you need the exact low-level move/action sequence for a request before execution.

Examples:

• "Move in a circle for 10 seconds." -> build a list like alternating right 250ms, forward 417ms, repeated
• "Move in a square for 5 seconds each side and then sit." -> build a list like forward 5000ms, six right-turn steps, repeated per side, then action sit

movement_pattern

Use movement_pattern for broader named patterns, especially triangle, figure-eight, and zigzag requests. Prefer circle_pattern for direct circle requests.

Parameters:

• pattern: one of square, triangle, circle, figure_eight, zigzag
• segment_duration_ms: duration for each side/leg for polygonal patterns, and total loop duration for circle
• turn_direction: left or right
• repeat_count: optional number of full repetitions
• final_actions: optional follow-up actions

Examples:

• "Do a figure eight and then dance." -> call movement_pattern with pattern: "figure_eight", final_actions: ["dance"]
• "Make a zigzag pattern twice." -> call movement_pattern with pattern: "zigzag", repeat_count: 2

execute_request

Use execute_request for natural-language requests that need interpretation before execution.

Examples:

• "Move in a circle for 20 seconds and then sit and wave." -> call execute_request
• "Do a square patrol and finish with a dance." -> call execute_request
• "Make a triangle, then stand." -> call execute_request

status

Use status to read the current MechDog or simulator state.

HTTP fallback:

curl -sS "http://${MECHDOG_IP:-localhost:3000}/status"

see

Use see to analyze the camera image with a vision model.

Parameters:

• question: optional analysis prompt
• provider: one of openai, nebius, anthropic; default openai

Environment

• Simulator default: MECHDOG_IP=localhost:3000
• Hardware example: MECHDOG_IP=192.168.1.100
• Vision example:
  • OPENAI_API_KEY=...
  • OPENAI_MODEL=gpt-5.4-mini