MobiAgent
Checkpoint at 25,000 training steps from the RoboCasa joint-training run. The shared vision-language backbone and six action experts are trained together.
Paper: From Execution to Evolution: A Dual-Loop Agentic System for Long-Horizon Mobile Manipulation
Code: kaiknower/MobiAgent
Checkpoint contents
| Path | Contents |
|---|---|
params/ |
Inference parameters in JAX / Orbax OCDBT format |
assets/ |
Expert names and normalization statistics |
train_state/ |
Training state for resuming with the matching training configuration |
_CHECKPOINT_METADATA |
Orbax checkpoint metadata |
checkpoint_info.json |
Model interface and checkpoint step |
checkpoint_manifest.json |
Checkpoint file paths and sizes |
The expert order is close, open, switch, manipulate, navigate, pnp. The policy uses three camera views, a 16-dimensional simulator state padded to 32 model dimensions, 12-dimensional simulator actions, and an action horizon of 50.
Download
Request access on this page and wait for approval, then authenticate:
hf auth login
Download the inference parameters and normalization assets:
hf download Liukaikai/MobiAgent \
--include 'params/**' --include 'assets/**' \
--include '_CHECKPOINT_METADATA' --include 'checkpoint_info.json' \
--local-dir checkpoints/robocasa-joint-25000
For the full checkpoint, including training state:
hf download Liukaikai/MobiAgent \
--local-dir checkpoints/robocasa-joint-25000
Use the MobiAgent OpenPI backend to load this Orbax checkpoint. See the
RoboCasa inference guide.
In the inference data recipe, point each expert's norm_path to its downloaded
assets/per_expert/<expert>/norm_stats.json. Preserve the expert order above.
uv run --project policy/openpi scripts/robocasa/serve.py \
--checkpoint /absolute/path/to/checkpoints/robocasa-joint-25000 --port 8000
The backend derives from OpenPI. Use the corresponding model implementation and normalization assets for inference and the matching optimizer and schedule configuration when resuming training.