vibe

implementation accompanying ViBe: Visual Behavior Adaptation for Perceptive Humanoid Whole-Body Control.

install

Requires uv and Git LFS; uv fetches Python itself.

git clone https://github.com/lok-i/vibe && cd vibe
uv venv --prompt vibe
source .venv/bin/activate
bash scripts/setup/sync_deps.sh     # Vibe + pinned deps
bash scripts/setup/sync_data.sh     # all | inhouse | omre | grail

data modes, dependencies, troubleshooting: docs/setup.md.

usage

# train
list-envs                                   # task ids
train Vibe-Repose-BigCubeFloor-ImgFeat-Ext --env.scene.num-envs 4096 \
  --agent.max-iterations 60000 --env.commands.motion.init-phase-anneal-iterations 50000 \
  --agent.wandb-project vibe

# play
play Vibe-Repose-BigCubeFloor-ImgFeat-Ext --agent release --viewer native
play Vibe-Uolm-ImgFeat-Ext --agent initial --viewer native

• checkpoints: lkrajan/vibe, one per task
• --agent release: fetched on first use into ~/.cache/vibe/releases
• all up front: bash scripts/setup/download_released_models.sh
• train commands, every task: docs/tasks.md

export ONNX

bash scripts/setup/sync_deps.sh --deploy           # + onnxruntime
export-agent   <task-id> --release                 # policy -> exports/agent/<task-id>/
export-agent   <task-id> --release --viewer native # + watch the two-world check
export-encoder --tag theia-tiny                    # vision backbone -> exports/enc/

for deployment support, see: docs/export.md.

test

ruff check src tests    # lint (also CI)
pytest tests/           # contracts over the synced data; no GPU, ~7 s

tasks

contribute

We welcome contributions, be it an item on our roadmap or new features. Feel free to open an issue/PR.

license

Code: BSD-3-Clause. Third-party models, data and the released checkpoints keep their own terms: THIRD_PARTY_NOTICES.md.