Downloads · 30 days
8
17% of all-time downloads
rocochallenge2025/rocochallengeIROS2026-pi05-roco-umi-29999
rocochallengeIROS2026-pi05-roco-umi-29999 is a robotics model from rocochallenge2025. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for openpi. The card lists the license as apache-2.0.
OpenPI JAX PI0.5 full-fine-tuning checkpoint for the RoCo UMI mobile-manipulation dataset. The checkpoint is intended for inference and real-robot integration through the OpenPI websocket server.
Downloads · 30 days
8
17% of all-time downloads
All-time downloads
48
Public
Repo size
24.9 GB
Likes
0
Public
Click a slice to open those files.
Other12.4 GB · 100%
From the Hugging Face model README
OpenPI JAX PI0.5 full-fine-tuning checkpoint for the RoCo UMI mobile-manipulation dataset. The checkpoint is intended for inference and real-robot integration through the OpenPI websocket server.
checkpoint/19999/params/
checkpoint/19999/assets/
deployment/serve_roco_umi.py
deployment/smoke_test.py
deployment/umi_client_template.py
deployment/roco_umi_config.py
deployment/umi_policy.py
config.json
The inference package contains params and the dataset normalization asset. The training optimizer state is intentionally excluded.
Download the complete inference bundle (about 12.4 GB):
export BUNDLE=$HOME/pi05-roco-umi-19999
uvx --from huggingface_hub hf download \
rocochallenge2025/rocochallengeIROS2026-pi05-roco-umi-19999 \
--local-dir "$BUNDLE"
Use the OpenPI source at a pinned revision compatible with this release, then install its dependencies:
git clone https://github.com/Physical-Intelligence/openpi.git
cd openpi
git checkout 215abfb
uv sync
Set the bundle path and run the smoke test:
export BUNDLE=/path/to/pi05-roco-umi-19999
export PYTHONPATH=$BUNDLE/deployment:$PWD/src
uv run python $BUNDLE/deployment/smoke_test.py \
--checkpoint $BUNDLE/checkpoint/19999
On the inference machine:
export BUNDLE=/path/to/pi05-roco-umi-19999
export PYTHONPATH=$BUNDLE/deployment:$PWD/src
uv run python $BUNDLE/deployment/serve_roco_umi.py \
--checkpoint $BUNDLE/checkpoint/19999 \
--host 0.0.0.0 \
--port 8000
The first inference performs JAX compilation and can be slow. Subsequent calls use the compiled policy.
The client sends three uint8 RGB images of shape 384x480x3, a float32 state vector of length 16, and an optional prompt.
State order:
left_x, left_y, left_z, left_qx, left_qy, left_qz, left_qw,
right_x, right_y, right_z, right_qx, right_qy, right_qz, right_qw,
left_gripper_command, right_gripper_command
The server returns an action chunk of shape (50, 16). Each action contains left and right Cartesian target pose (x,y,z,qx,qy,qz,qw) followed by the two gripper commands. Quaternion outputs are normalized before they reach the client.
The model is not a generic motor controller. The hardware adapter must implement camera calibration, coordinate-frame conversion, Cartesian tracking, gripper mapping, workspace limits, collision checks, watchdogs, and emergency stop behavior. Start with low speed and a lifted, collision-free test pose.
See deployment/umi_client_template.py for the adapter interface.
This package follows the license terms of the upstream OpenPI code and the released dataset. Check the upstream and dataset licenses before redistribution. This model has passed a native load/output smoke test; that is not a safety certification or a guarantee of task success on a physical robot.