Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/zju3dv/habitat-gsnpx agentmods add skills/zju3dv/habitat-gs/habitat-gs-trainWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/zju3dv/habitat-gs/habitat-gs-train)<a href="https://agentmods.dev/skills/zju3dv/habitat-gs/habitat-gs-train"><img src="https://agentmods.dev/badge/skills/zju3dv/habitat-gs/habitat-gs-train/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/zju3dv/habitat-gs/habitat-gs-train"><img src="https://agentmods.dev/badge/skills/zju3dv/habitat-gs/habitat-gs-train.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00114 | $0.02016 |
| Opus 5 | $0.00057 | $0.01008 |
| Sonnet 5 | $0.00023 | $0.00403 |
| Haiku 4.5 | $0.00011 | $0.00202 |
Grade A, and why
habitat-gs-train scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 13d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 88 lines — stays where its author put it; the contents beside it link to each section on GitHub.
habitat-gs-train
Train and evaluate navigation policies on photo-realistic 3D Gaussian Splatting scenes in habitat-gs. Five tasks are supported, each with a one-click generate → train → evaluate pipeline driven by scripts under scripts_gs/.
IMPORTANT — run from the repo root. Every command below assumes the current directory is the habitat-gs/ project root and that the right conda env is active. The scripts cd to the project root themselves, but paths like data/scene_datasets/gs_scenes/... are repo-relative. Never invent flags — the exact flags each script accepts are in references/.
Pick the task first
| Task | Kind | Conda env | Backbone | Reference |
|---|---|---|---|---|
| PointNav | RL (Habitat-Lab + DDPPO) | habitat-gs |
PointNavResNet-50 + LSTM | references/task-pointnav.md |
| ImageNav | RL (Habitat-Lab + DDPPO) | habitat-gs |
PointNavResNet-50 + LSTM | references/task-imagenav.md |
| ObjectNav | RL (Habitat-Lab + DDPPO) | habitat-gs |
PointNavResNet-50 + LSTM | references/task-objectnav.md |
| StreamVLN | VLN (VLM supervised fine-tune) | habitat-gs-streamvln |
LLaVA-Video-7B-Qwen2 + SigLIP | references/task-streamvln.md |
| Uni-NaVid | VLN (VLM supervised fine-tune) | habitat-gs-uni-navid |
Vicuna-7B + EVA-ViT-G | references/task-uninavid.md |
RL tasks (PointNav / ImageNav / ObjectNav) are the lightweight, fast path — they run in the base habitat-gs env and need no external repo. Start here for "quickly train and evaluate a navigation policy."
VLN tasks (StreamVLN / Uni-NaVid) are heavy: each needs ≥80 GB VRAM/GPU (StreamVLN has a 24 GB LoRA mode), a separate cloned conda env, and an external repo cloned as a sibling of habitat-gs/. Only go here when the user explicitly wants instruction-following VLN.
The universal flow (all tasks)
- Prerequisites — env + Habitat-Lab + GS data. Read
references/prerequisites.mdand verify before anything else. Skipping this is the #1 cause of failures. - Generate episodes / trajectories — produces the dataset the policy trains/evals on. The released dataset already ships episodes, so this is OPTIONAL for the standard scenes (the train/eval scripts abort with a clear message if the data is missing).
- Train —
bash scripts_gs/train_<task>.sh --output <dir> [options]. - Evaluate —
bash scripts_gs/eval_<task>.sh --ckpt <ckpt> [options]; read where metrics land inreferences/outputs-and-metrics.md.
What ships with it
10 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- references/data-layout.md 3.4 KB
- references/outputs-and-metrics.md 1.9 KB
- references/prerequisites.md 2.8 KB
- references/task-imagenav.md 2.6 KB
- references/task-objectnav.md 3.1 KB
- references/task-pointnav.md 3.7 KB
- references/task-streamvln.md 3.5 KB
- references/task-uninavid.md 2.9 KB
- references/training-and-finetuning.md 4.2 KB
- references/troubleshooting.md 3.1 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 13d ago First seen · 88 lines · 114 tokens per session scan A cde44ac59d03
habitat-gs-train is a skill published in the GitHub repository zju3dv/habitat-gs (295 stars, last pushed 1mo ago), licensed MIT. It adds 114 tokens to every session and 2,016 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
mjlab-to-mjswan
Port an mjlab task to mjswan: turn a local or remote GitHub repo that registers mjlab tasks into a browser app. Use when the user wants to run, visualize, or deploy an mjlab task (theirs or a third party's) in the browser with mjswan.
configuring-vision
The user wants to connect an LLM or vision provider, already has an API key, asks "can I use OpenAI/Anthropic/Gemini/OpenRouter", wants local Ollama, or needs different cheap and strong models. Use this to configure provider-neutral visual understanding without tying Watch Skill to one agent or model vendor.
tao-validate-recipe-transfer
Port a published computer vision paper's official code and training recipe onto a customer's own dataset, or diagnose why such a transfer produced bad numbers. Use this whenever someone wants to reproduce a CV paper, run a paper's repo on their own images, fine-tune a published…
authoring-pipelines
Use when the user wants a bagel data pipeline — reducing a log to windows around events ("keep 30s around every hard brake"), recurring snippets/GIFs/exports, or any tasks-and-gates automation over robot data.
segment-anything-model
Foundation model for image segmentation with zero-shot transfer. Use when you need to segment any object in images using points, boxes, or masks as prompts, or automatically generate all object masks in an image.
perceiving-object-parts
Hierarchical perception for subpart targeting. Detects a parent object first (DINO+VLM), crops the camera image to the parent's bounding box, then detects and segments the named subpart inside the crop (DINO + SAM3), and uncrops + fuses depth to a world-frame OBB/mask/cloud — plus the parent object's OBB and cloud for…