openclaw-vln-planner（OpenClaw VLN planner 搜索）

OpenClaw VLN Planner

Use this skill when the user wants a robot to follow a natural-language navigation instruction from visual observations.

This skill is a high-level navigation planner. It does not produce motor, joint, torque, or trajectory control. It only produces one structured mid-level navigation action at a time.

When this skill triggers

Trigger this skill when the task includes one or more of the following:

• Vision-language navigation (VLN)
• Robot next-step planning from camera images
• Closed-loop navigation with replanning after each observation
• Converting a current frame plus historical frames into a single next navigation action
• Sending current + history images to an OpenAI-compatible multimodal gateway for action prediction

Required inputs

The planner expects:

• user_instruction: natural-language navigation instruction
• current_frame: exactly one current image
• history_frames: zero or more previous images in temporal order

Optional inputs:

• robot_state: heading, speed, pose estimate, execution feedback, etc.
• safety_flags: blocked, collision_risk, lost, target_reached, low_visibility, etc.
• config_path: path to the runtime config file

Output contract

Output must be pure JSON only. Do not prepend or append prose.

Allowed action types only:

• MOVE_FORWARD
• TURN_LEFT
• TURN_RIGHT
• STOP

Expected JSON shape:

{
  "next_action": {
    "type": "MOVE_FORWARD",
    "value": 75,
    "unit": "cm"
  },
  "task_status": "in_progress",
  "confidence": 0.87,
  "notes": "continue along the hallway"
}

Completion example:

{
  "next_action": {
    "type": "STOP"
  },
  "task_status": "completed",
  "confidence": 0.93,
  "notes": "goal reached"
}

Core rules

1. Plan only the next action.
2. Never output a full route.
3. Replan after each execution step.
4. If uncertain, unsafe, blocked, unable to parse, or visually ambiguous, output STOP.
5. Enforce action bounds:

- MOVE_FORWARD: 10-150 cm - TURN_LEFT: 5-90 deg - TURN_RIGHT: 5-90 deg - STOP: no value/unit required

1. If safety_flags.target_reached == true, output STOP with task_status = completed.
2. If blocked, collision_risk, lost, or severe uncertainty is present, prefer STOP.

Runtime configuration

Before running, load a YAML config file such as config/vln-config.yaml.

The config should define:

• subscribed or logical input topics / channels for current frame and history frame collection
• optional robot state and safety flag sources
• OpenAI-compatible multimodal gateway settings: base_url, api_key, model_id
• planner behavior such as confidence threshold and safety fallback
• executor bridge mode (default: Python function bridge)

Read references/navigation-schema.md for the expected config structure.

Internal module design

1) context builder

Build a model input payload from:

• user instruction
• historical observations
• current observation
• optional robot state
• optional safety flags

The prompt must explicitly separate:

• historical observations
• current observation
• user instruction

2) action planner

Call an OpenAI-compatible multimodal gateway with:

• one current image
• historical images
• planner prompt
• optional structured context

The model should be asked to return pure JSON for exactly one next action.

3) action parser

Parse the model result as JSON.

If parsing fails:

• try safe extraction of the first JSON object
• if still invalid, fall back to STOP

4) action validator

Validate:

• action type is one of the four allowed values
• distance and angle ranges are legal
• unit matches action type
• confidence is numeric if present
• task_status is one of in_progress, completed, failed

Any invalid output falls back to STOP.

5) executor bridge

Forward the validated mid-level action to a separate execution layer.

Reserved Python bridge interface:

• execute_move_forward(distance_cm)
• execute_turn_left(angle_deg)
• execute_turn_right(angle_deg)
• execute_stop()
• get_robot_state()
• get_safety_flags()

Do not hardcode a robot SDK into the planner logic.

6) replanning loop

Use the planner in a closed loop:

1. gather current frame + history frames
2. gather optional robot state / safety flags
3. call multimodal planner
4. parse and validate JSON action
5. execute through bridge
6. observe again
7. repeat until task_status = completed or forced stop

7) safety fallback

Always stop on:

• parse failure
• invalid action
• confidence below threshold
• blocked / collision risk / lost / target reached
• missing visual evidence for safe motion

Prompt template

Use this prompt pattern:

You are a robot navigation planner.
You will receive:
1. historical observations
2. current observation
3. a user instruction
4. optional robot state and safety flags

Your job is to decide the robot's next single mid-level navigation action.
You may output only one of these actions:
- MOVE_FORWARD with distance in cm
- TURN_LEFT with angle in deg
- TURN_RIGHT with angle in deg
- STOP

Rules:
- Plan only the next step, not the whole route.
- If the goal has been reached, output STOP.
- If you are uncertain, the scene is unclear, or there is any safety risk, output STOP.
- MOVE_FORWARD must be 10-150 cm.
- TURN_LEFT and TURN_RIGHT must be 5-90 deg.
- Output pure JSON only, with no extra explanation.

Example user requests

• "Go down the hallway and stop at the blue door."
• "Move to the kitchen entrance."
• "Find the end of the corridor and stop."
• "Turn right at the next intersection and continue."

Failure handling

If anything is wrong with the output, return:

{
  "next_action": {
    "type": "STOP"
  },
  "task_status": "failed",
  "confidence": 0.0,
  "notes": "fallback_stop"
}

Bundled resources

• references/navigation-schema.md: schema, bounds, safety fallback, examples, config contract
• scripts/vln_bridge.py: example OpenAI-compatible multimodal planner + Python executor bridge
• scripts/requirements.txt: Python dependencies
• config/vln-config.yaml: runtime config template