Skip to content
Other

DriveSTA

trajectory bank and checkpoints for autonomous driving

2 min read

This repository contains the full release of data and trained models from the DriveSTA project for autonomous vehicle motion planning. Inside, you will find 308,119 driving scenes with precomputed trajectory options and four weight files (checkpoints) in Base and Scale configurations across two training stages (Stage 1 and Stage 2). Unlike standard datasets that only store raw logs or images, this release provides pre-scored sets of 3,072 candidate trajectories for every scene. It is aimed at researchers and engineers working with the NAVSIM environment and vehicle route planning tasks.

Key features

  • Large-scale trajectory bank: 103,288 real-world scenes from Navtrainval and 204,831 synthetic scenes from SimScale, totaling around 135 GiB compressed.
  • Four ready-to-use checkpoints: the files base_s1.ckpt, base_s2.ckpt, scale_s1.ckpt, and scale_s2.ckpt, preserving the original training results bit-for-bit.
  • Detailed data structure: each scene includes 3,072 trajectories in [8, 3] format (x, y coordinates and heading angle in the vehicle coordinate frame at 0.5-second intervals), along with their scores and metadata.
  • Easy handling of large archives: the data is split into independent .tar.gz shards with SHA-256 manifests and scripts for selectively recovering corrupted fragments.

What you can do

  • Train and test planning models: use the prebuilt trajectory banks and fixed training splits (covering 232,672 scenes) to reproduce experiments in NAVSIM.
  • Run inference on pretrained weights: plug the Base or Scale checkpoints into the core model code from the project's GitHub repository to evaluate trajectory prediction accuracy.
  • Read arrays quickly via mmap: work with the numerical banks directly from disk using NumPy without loading the entire dataset into RAM.

Example

hf download Rtwotwo/DriveSTA --local-dir DriveSTA
cd DriveSTA
sha256sum -c docs/release/SHA256SUMS
python scripts/extract_release.py weights/Trajectories/*/*.tar.gz --destination . --workers 4
python scripts/verify_trajectories.py --root . --workers 4
Read moreDetails
Total size of binary files is 146.44 GiB, plus around 170 MiB of manifests and checksums.
Unpacking and verification require Python 3.9+ (standard library), while reading the arrays requires NumPy.
Images, maps, raw logs, and feature caches are not included in the repository and must be downloaded separately.
The release code and metadata are distributed under the Apache-2.0 license.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 5.3 GB · ready files6.3 GB
  • FP8 / INT8recommendedweights 2.7 GB3.4 GB
  • 4-bitweights 1.5 GB2.1 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 5060 Ti 8GB

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.

    # Rtwotwo/DriveSTA
    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 Rtwotwo/DriveSTA --local-dir ~/models/DriveSTA
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to DriveSTA
  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

New posts about versions of DriveSTA (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

Rtwotwo/DriveSTA
Author
Rtwotwo
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