ViBe release models
Public release checkpoints for ViBe — Visual Behavior adaptation: vision adapters over a frozen SONIC base on the Unitree G1, one per task family. These are the behavior-validated checkpoints selected from the corresponding private training runs. The W&B project is not required to download or play them.
Play immediately
After installing ViBe and its pinned dependencies:
# Optional: download and verify every checkpoint up front.
bash scripts/setup/download_released_models.sh
# Or let play download only the selected checkpoint on first use.
play Vibe-Repose-BigCubeFloor-ImgFeat-Ext --agent release --viewer native
play Vibe-PerLoco-Grail-ImgFeat-Ext --agent release --viewer native
play Vibe-PerLoco-OmRe-ImgFeat-Ext --agent release --viewer native --num-envs 27
play Vibe-Uolm-ImgFeat-Ext --agent release --viewer native
play Vibe-Dodge-ConeFast-ImgFeat-Ext --agent release --viewer native
Downloads are cached by release tag under ~/.cache/vibe/releases by default and are verified
against the SHA-256 values in release.json. Set VIBE_RELEASE_ROOT to move the cache.
Contents
| task | checkpoint | training iteration |
|---|---|---|
Vibe-Dodge-ConeFast-ImgFeat-Ext |
Vibe-Dodge-ConeFast-ImgFeat-Ext/checkpoint.pt |
14,999 |
Vibe-PerLoco-Grail-ImgFeat-Ext |
Vibe-PerLoco-Grail-ImgFeat-Ext/checkpoint.pt |
14,999 |
Vibe-PerLoco-OmRe-ImgFeat-Ext |
Vibe-PerLoco-OmRe-ImgFeat-Ext/checkpoint.pt |
14,999 |
Vibe-Repose-BigCubeFloor-ImgFeat-Ext |
Vibe-Repose-BigCubeFloor-ImgFeat-Ext/checkpoint.pt |
59,999 |
Vibe-Uolm-ImgFeat-Ext |
Vibe-Uolm-ImgFeat-Ext/checkpoint.pt |
24,999 |
Each .pt file is the complete RSL-RL training checkpoint: actor (frozen SONIC base, LoRA
adapter, attention extractor), critic, optimizer state, and iteration. ViBe loads the actor for
inference. The four deployment exports below are also available; export-agent <task>
produces new exports from your own checkpoints.
The frozen vision backbone is separate from the .pt checkpoints. Training fetches it from
its upstream repository; deployment uses the exported encoder in encoders/, under its own license. The PerLoco, UOLM and Repose tasks need
their motion/terrain datasets; scripts/setup/sync_data.sh in the ViBe repository stages them.
The privileged twins of these tasks are published separately as
lkrajan/orcs.
Deploy on the G1
Use vibe_ws for the simulator and ROS 2 commands.
Keep each ONNX file beside its matching .manifest.json.
| sample | policy + manifest, beside checkpoint.pt |
|---|---|
| walk | Vibe-PerLoco-Grail-ImgFeat-Ext/model_14999.{onnx,manifest.json} |
| parkour | Vibe-PerLoco-OmRe-ImgFeat-Ext/model_14999.{onnx,manifest.json} |
| repose | Vibe-Repose-BigCubeFloor-ImgFeat-Ext/model_59999.{onnx,manifest.json} |
| uolm | Vibe-Uolm-ImgFeat-Ext/model_24999.{onnx,manifest.json} |
All four use encoders/theia-tiny.onnx and encoders/theia-tiny.manifest.json
(112×63 BGR uint8 input). deployment.json records checksums,
checkpoint lineage, historical exporter revisions and validation limits.
The encoder is a modified ONNX export of frozen Theia.
Dodge's checkpoint remains available; it is outside this four-sample deployment bundle.
Deployment downloads use the immutable HF commit recorded in vibe_ws/artifacts.lock.json.
The original v0.1.0 tag remains the checkpoint-only release.
Versioning and provenance
Use the immutable v0.1.0 revision for reproducible training-checkpoint downloads. release.json records the exact
file paths, byte sizes, checksums, and the validated Mocke/RSL-RL/ORCS revisions. Each task folder
also contains a small provenance.json; private W&B run IDs are retained only as lineage
identifiers.
License
The checkpoints incorporate SONIC model weights and are distributed under the NVIDIA Open Model License. Licensed by NVIDIA Corporation under the NVIDIA Open Model License.
ViBe-authored code and documentation are BSD-3-Clause. External models, datasets, and assets —
the frozen vision backbones included — retain their respective terms; see NOTICE and
THIRD_PARTY_NOTICES.md.
These are special-purpose robotics control models. Validate them in simulation before any use on hardware.