OpenClaw VLN Planner

Plan the next high-level navigation step for a robot from a user navigation instruction, one current image, and a sequence of historical images. Use when the...

This skill appears to do what it says: it will read local camera frame files, base64-encode them, and POST them (with the model API key from a YAML config) to whatever OpenAI-compatible gateway you configure. Before installing: (1) confirm you trust the gateway endpoint and operator because camera images and any scene data will be transmitted; (2) store the API key securely (the example uses a config file rather than env vars) and update the manifest if you need policy/audit visibility; (3) keep executor in dry_run while testing and review/replace the placeholder execute_* functions so the planner cannot command hardware until you've integrated a vetted execution bridge; (4) if you require stricter telemetry controls, inspect/modify image_to_data_url and build_messages to avoid sending raw images or to anonymize them. The small manifest omission (no declared required config path/credential) is not malicious but worth correcting for clarity and safety.

Type: OpenClaw Skill Name: openclaw-vln-planner Version: 1.0.0 The skill bundle is a legitimate vision-language navigation (VLN) planner designed to convert camera observations and user instructions into robot movement commands. The Python bridge (`scripts/vln_bridge.py`) implements standard multimodal LLM integration using the `requests` library and includes robust validation logic, such as movement distance/angle bounds and safety fallbacks. No evidence of data exfiltration, malicious execution, or harmful prompt injection was found; the file access and network activity are strictly aligned with the stated purpose of processing image frames and communicating with an AI gateway.