Skip to main content

Checkpoint

Your computer
Sample output

Inputs and output

Deploy and predict

Deploy the checkpoint URI directly. Orchestral resolves, imports and prepares it before bringing up the serving deployment.
your computer
Sample output
robot computer

Replan with actions already queued

Pass queued rows from an earlier prediction as committed_actions. Supply ordinary float32 actions in the same format as result.actions; Orchestral handles the selected pi0.5 policy’s normalization and action transforms. The returned prefix is preserved exactly, while the remaining rows are generated conditioned on it. The prefix can be any length up to the checkpoint’s action horizon.
robot computer

Select the inference policy

OpenPI requires policy_config during inspection and deployment. Use the registered upstream TrainConfig name that matches training:
your computer
Orchestral runs that config’s input transforms, normalization, model transforms, inference, unnormalization, and output transforms. inspect reports the request names and shapes for that policy. For example, pi05_aloha accepts images/cam_high, images/cam_left_wrist, images/cam_right_wrist, and state. Supply those exact names in session.predict(inputs=...). The bundle supports native JAX pi0.5 checkpoints. Policy names come from its pinned OpenPI version. An unknown name or a missing policy_config raises an error. Custom Python training configs require their input and output transforms to run in your client.

Client-owned transforms

Use policy_config=none when your client applies the training-specific input and output transforms. The YAM example above uses this setting. For a custom checkpoint, also supply its trained action horizon:
your computer
The server accepts base_0_rgb, left_wrist_0_rgb, right_wrist_0_rgb, and state. It applies saved-stat normalization, the OpenPI model transforms, inference, and action unnormalization. Returned actions have model padding removed. Your client applies its training-specific state conversion and converts returned actions to robot commands, including any gripper conversion, delta-to-absolute conversion, and clipping. All three camera slots are required. Supply RGB uint8 arrays with shape (224, 224, 3). Use the views and state order from training. The YAM checkpoint has a recorded horizon of 15; other checkpoints require action_horizon in this mode.

Normalization assets

A named policy supplies its upstream asset_id. With policy_config=none, deployment selects a sole bundled assets/<asset_id>/norm_stats.json. Select a particular asset with asset_id:
your computer
Orchestral first checks assets/trossen/norm_stats.json in your checkpoint. If that file is absent, it uses the selected upstream config’s assets_dir when one is declared. Otherwise, it searches the official asset roots exported from the pinned OpenPI configs, including base, DROID, and Polaris assets. Multiple matches raise an error listing the choices. An unknown ID raises an error listing available upstream assets. Use the statistics that match your training run’s state and action definitions. Deployment records the asset ID, source, and SHA-256 checksum. Upstream sources also include the GCS object generation. Workers verify the selected bytes when loading the model; a missing or changed imported asset fails loading. Bundled assets remain tied to the imported checkpoint. Inspect the resolved asset and inference settings before deployment:
your computer
The JSON response includes resolved_checkpoint_options. Deployment saves these options and passes them to deployment workers. Checkpoints use OpenPI’s existing parameter and statistics files.

Inference settings

A named policy supplies the model settings used by its upstream TrainConfig. With policy_config=none, the following settings apply: Pass multiple options after one flag, for example --options policy_config=none action_horizon=15 max_token_len=128. Your client controls request cadence and action execution timing.