hjyeee/Follow-Bench-2.0-packages

FollowBench 2.0 release packages

The custom CARLA build and the model weights used by the FollowBench 2.0 benchmark. The benchmark code lives in the FollowBench 2.0 GitHub repository at https://github.com/MedlarTea/follow-bench-2.0; its installer downloads, verifies and unpacks these files for you.

This note is not legal advice. The license that applies to each model is the text shipped with it.

Install (recommended)

From a clone of the FollowBench 2.0 repository, inside its conda env:

bash tools/install/install_assets.sh --download --install-wheel core   # CARLA + perception weights
bash tools/install/install_assets.sh --download all                    # + OA-VAT + TrackVLA
python tools/install/check_install.py all                              # exit 0 = complete

The installer resumes interrupted downloads, checks each zip's size and SHA256, unpacks it into the repository, fetches the AGPL-3.0 YOLO weights from their publishers, and finally checks every unpacked file against the per-file SHA256 list in the repository (tools/install/files.tsv). Set HF_ENDPOINT=https://hf-mirror.com if huggingface.co is not reachable.

Manual download

pip install -U huggingface_hub
hf download hjyeee/Follow-Bench-2.0-packages --repo-type dataset --local-dir downloads
cd downloads && sha256sum -c SHA256SUMS.txt

Then either run bash tools/install/install_assets.sh all from the repository root (it picks up the zips in downloads/), or unzip by hand: unzip <zip> -d <repo> for the three weight zips and unzip followbench2_carla_multi_robots.zip -d <repo> for CARLA.

Contents

File	Size	Unpacks to	Contains	License
followbench2_carla_multi_robots.zip	5.8 GB	CARLA_multi_robots/ (14.2 GB)	CARLA 0.9.16 build with Scout / Dog / G1 robots (Linux), its Python client wheel	CARLA_multi_robots/LICENSE
followbench2_weights.zip	3.7 GB	data/weights/ (4.0 GB)	R18 ReID ckpt.t7, trajectory predictors sgan.pt csgan.pt csgan2.pt, 10 KPR checkpoints	KPR: Hippocratic License 3.0
followbench2_oa_vat.zip	0.6 GB	data/oa-vat/ (0.7 GB)	ORTrack ORTrack_ep0300.pth.tar, MobileCLIP-B-LT mobileclip_blt.ts	ORTrack: MIT. MobileCLIP: Apple ML Research Model License (non-commercial)
followbench2_trackvla.zip	16.3 GB	data/trackvla/ (19.8 GB)	TrackVLA checkpoint, Qwen3-4B, SigLIP so400m-patch14-384	Qwen3-4B, SigLIP: Apache-2.0

SHA256SUMS.txt lists the SHA256 of the four zips.

Not included: download these yourself

These weights are needed by some baselines but are not redistributed here. check_install.py reports which are missing and checks the SHA256 of the ones you add.

File (relative to the repository)	Needed for	Why not here	How to get it	SHA256
data/weights/yolo/yolo11s.pt, yolo11n.pt	all perception baselines	AGPL-3.0	fetched automatically by install_assets.sh from Ultralytics/YOLO11	85a76fe8…502d5, 0ebbc80d…644ee1
data/oa-vat/yolo/yoloe-11l-seg.pt	OA-VAT	AGPL-3.0	fetched automatically from jameslahm/yoloe	a993fb0f…3699f
data/trackvla/facebook/dinov3-vits16-pretrain-lvd1689m/ (config.json, preprocessor_config.json, model.safetensors)	TrackVLA	gated by Meta (DINOv3 License)	accept the license at https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m, then hf download facebook/dinov3-vits16-pretrain-lvd1689m --local-dir data/trackvla/facebook/dinov3-vits16-pretrain-lvd1689m	model.safetensors: 4610ad75…c4f91d
data/oa-vat/dinov3/dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth	OA-VAT	gated by Meta (DINOv3 License)	request access at https://ai.meta.com/resources/models-and-libraries/dinov3-downloads/, download with the link from the approval email	73cec8be…081a7c
data/oa-vat/dinov3/dinov3_vits16_pretrain_lvd1689m-08c60483.pth	OA-VAT with --oa-vat-dinov3 dinov3_vits16 only	same	same	starts with 08c60483

The full SHA256 values are in tools/install/files.tsv of the repository.

License notes for included weights

• MobileCLIP-B-LT (data/oa-vat/mobileclip_blt.ts): Apple Machine Learning Research Model, licensed under the Apple Machine Learning Research Model License Agreement (data/oa-vat/LICENSE_MODELS, MOBILECLIP_NOTICE.txt). Use is limited to non-commercial scientific research and academic development. Upstream: https://github.com/apple/ml-mobileclip.
• KPR (data/weights/reid_kpr/): Hippocratic License 3.0 (HL3-LAW-MEDIA-MIL-SOC-SV), full text in data/weights/reid_kpr/LICENSE. Redistribution is allowed, including commercial use, and derivative works must keep this license. The ethical-use conditions restrict mass surveillance, military activities, providing goods or services to law enforcement, and broadcasting extreme violence, among other things. Upstream: https://github.com/VlSomers/keypoint_promptable_reidentification.
• ORTrack: MIT, Copyright (c) 2025 You Wu, data/oa-vat/ortrack/LICENSE. https://github.com/wuyou3474/ORTrack
• SigLIP so400m-patch14-384: Apache-2.0. The Hugging Face snapshot has no LICENSE file, so the zip adds the Apache-2.0 text from https://github.com/google-research/big_vision.
• Qwen3-4B: Apache-2.0, data/trackvla/LLM_hf/qwen3-4b/LICENSE.
• CARLA: the CARLA code is MIT (CARLA_multi_robots/LICENSE). Upstream CARLA states that its assets are CC-BY, and the packaged binary contains Unreal Engine 4 components covered by the Unreal Engine EULA. See https://github.com/carla-simulator/carla#licenses.