--- license: apache-2.0 tags: - autonomous-driving - trajectory-planning - navsim - pytorch --- # DriveSTA: trajectory banks and trained checkpoints [Code](https://github.com/Rtwotwo/DriveSTA) · [Project page](https://rtwotwo.github.io/DriveSTA/) This repository provides the complete released **308,119-scene** trajectory collection and four trained **Base/Scale × Stage 1/Stage 2** checkpoints. The binary assets total **146.44 GiB**, plus approximately 170 MiB of scene inventories and file checksums. Each trajectory shard is an independent `.tar.gz`; model checkpoints are directly usable `.ckpt` files. Archive filenames remain stable so existing downloads and integrity manifests continue to match after the repository rename. | Collection | Scenes | Shards | Compressed size | | --- | ---: | ---: | ---: | | Navtrainval | 103,288 | 16 | 32.28 GiB | | SimScale | 204,831 | 32 | 102.78 GiB | | Shared metadata and immutable training indexes | — | 1 | 0.27 GiB | ```text weights/ Trajectories/ Navtrainval/ # 16 independent archives SimScale/ # 32 independent archives Metadata/ # drivetraj_metadata.tar.gz Checkpoints/ base_s1.ckpt base_s2.ckpt scale_s1.ckpt scale_s2.ckpt manifests/ scene_to_shard.csv # one row per scene shard_summary.csv # one row per archive index.json # totals, locations and SHA-256 values shards/ # per-file sizes and SHA-256 for each archive ``` ## Download, verify and recover See the [complete download and recovery guide](docs/DOWNLOAD_VERIFY.md), including a Chinese quick reference. The [scene-to-shard CSV](manifests/scene_to_shard.csv) maps every scene token to its dataset, split, shard and extraction directory. The [compact shard inventory](manifests/shard_summary.csv) lists all archive sizes, scene counts and SHA-256 values. ```bash 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 ``` Locate a scene, or repair only a damaged shard's extracted files: ```bash python scripts/verify_trajectories.py --root . --dataset navtrainval --token SCENE_TOKEN --locate python scripts/verify_trajectories.py --root . --shard drivetraj_navtrainval_001-of-016.tar.gz --repair ``` The repair helper verifies the archive before restoring missing or incorrect files and retains files that already pass their checks. A broken archive needs only that single shard to be downloaded again. The helpers require Python 3.9+ and use the standard library; reading trajectory arrays also requires NumPy. ## Trajectory contents Each scene includes 3,072 full candidate trajectories, scores and construction metadata, plus its available high-score subset. Trajectories have shape `[8, 3]`: ego-frame x, y and heading, sampled every 0.5 seconds. Numerical banks can be read with NumPy mmap using `scripts/read_trajectories.py`. Images, original logs, maps, feature/evaluation caches and EPDMS-mined trajectory banks are separate inputs and are not bundled here. The original high-bank shortfalls (25 real and 8,514 synthetic scenes) are preserved; their status markers describe mining outcomes, not download corruption. The fixed Scale training recipe uses 232,672 scenes (103,263 real + 129,409 synthetic). This differs from the complete released 308,119-scene collection; immutable training indexes are supplied in the metadata archive. ## Checkpoints and provenance The four checkpoint files preserve the original training outputs byte for byte. The full NAVSIM/model/mining implementation and external pretrained backbones are maintained separately from this asset repository. Original code and release metadata use Apache-2.0; third-party datasets and models retain their own terms. Archive and extracted-file SHA-256 checks, 503 tests, configuration checks and real-scene damage/recovery trials passed. See the [validation summary](docs/release/validation_summary.json), [asset inventory](docs/release/files.json) and [archive inventory](docs/release/archives.json). Full training and full benchmark evaluation were not rerun for this packaging step.