Skip to content
Other

CAT

Terrain traversability estimation for different robots from a single RGB-D frame

2 min readupdated

Developed by researchers at the University of São Paulo and presented at IROS 2026, CAT generates a dense traversability map from a single RGB-D image (a color frame paired with a depth map). Its standout feature is robot-aware prediction: the output adapts to the capabilities of the specific platform traversing the terrain, meaning an obstacle passable for a tracked vehicle might be flagged as impassable for a small wheeled rover. Under the hood, it relies on frozen backbones (feature-extraction networks) facebook/dinov3-vitb16-pretrain-lvd1689m and CIDAS/clipseg-rd64-refined. It is geared toward developers working on autonomous navigation and mobile robotics in unstructured, off-road environments.

Key features

  • Robot-specific adaptation: traversability predictions account for the capabilities and scores of the selected platform.
  • Single-frame inference: generates a map from a single RGB and depth image pair without needing a sequence of frames.
  • DINOv3 and CLIPSeg backbone: builds on frozen visual and multimodal features from pretrained models.
  • Bundled sensor statistics: ships with depth_stats.json containing per-sensor depth statistics alongside the best_model.pth weights.

What you can do

  • Off-road path planning: integrate the model into an autonomous robot's navigation stack to select safe trajectories across uneven terrain.
  • Cross-platform traversability comparison: run the same RGB-D scenes using different robot capability profiles (via scores.yaml) to see where each vehicle can pass.
  • Test on the NaviTrace dataset: run the ready-to-use cat_demo.py script on sample data from the project repository to evaluate segmentation quality.

Example

uv run --locked --extra cu128 hf download CAT-IROS-2026/cat best_model.pth depth_stats.json --local-dir checkpoints/cat

uv run --locked --extra cu128 python cat_demo.py \
  --checkpoint checkpoints/cat/best_model.pth \
  --rgb examples/navitrace/*/rgb.png \
  --depth examples/navitrace/*/depth.png \
  --scores examples/navitrace/scores.yaml \
  --sensor ts
Read moreDetails
Base models: facebook/dinov3-vitb16-pretrain-lvd1689m and CIDAS/clipseg-rd64-refined (downloaded automatically on first run).
Release contents: best_model.pth checkpoint and depth_stats.json depth statistics file.
Licenses: the checkpoint is distributed under the DINOv3 License (as it was trained on DINOv3 features), and the code is under the MIT License.
VRAM requirements and benchmark figures are not listed in the model card (see arXiv:2607.20679 for details).
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # CAT-IROS-2026/cat
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download CAT-IROS-2026/cat --local-dir ~/models/cat
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to cat
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. dinov3-vit7b16-pretrain-lvd1689mfamily root
  2. dinov3-vitb16-pretrain-lvd1689m
  3. cat

New posts about versions of dinov3-vit7b16-pretrain-lvd1689m (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 6All on Hugging Face

About the model

CAT-IROS-2026/cat
Author
CAT-IROS-2026
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments