> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orchestrallabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Isaac 0.5

> Inspect, deploy, and prepare fine-tuned Isaac 0.5 checkpoints.

## Checkpoint format for fine-tunes

Isaac 0.5 is imported as a complete Hugging Face snapshot because the policy weights, Qwen
language model, image processor, and LeRobot policy processors use relative paths. Do not upload
only the policy directory or only the changed weight shards. Preserve the release's root-relative
layout and update the metadata to describe the fine-tuned weights.

The snapshot must include:

| Path | Purpose |
| - | - |
| `config.json` | Isaac recipe and relative paths (`hf_model_path: ".."`, `fast_processor_path: "../fast_processor_pinned"`) |
| `model.safetensors.index.json` and every referenced shard | Qwen model weights |
| `lerobot_policy/config.json` | Policy dimensions, cameras, chunk, observation history, and target rate |
| `lerobot_policy/policy_preprocessor.json` and `policy_postprocessor.json` | Matching input and output processors; the preprocessor includes the enabled Isaac Harmony pack |
| `fast_processor_pinned/processor_config.json` and `tokenizer_config.json` | Pinned image and language preprocessing configuration |
| Remaining files referenced by those configs | Policy weights, tokenizer assets, and processor state |

The camera order, image size, action and proprioception dimensions, chunk size, number of
observation and action steps, dtype, inference backend, and target rate must agree between the
Isaac recipe and its processor. Servo validates that agreement during inspection/import. Keep the
whole snapshot in one repository revision so relative references cannot resolve to mismatched files.

## Inspect the input contract

Inspection reads metadata and small config files before deployment; it does not download the model
weights. For the public revision, the input contract has two image arrays named from its camera
order (`observation.images.image` and `observation.images.wrist_image`) and an eight-value
`observation.state` array. Check your own fine-tune's result because its dimensions may change.

Run the SDK inspection to print the checkpoint's input keys and shapes.

```python inspect_checkpoint.py theme={null}
import servo

sv = servo.Servo()
inspection = sv.models.inspect_checkpoint(
    "hf://PerceptronAI/Isaac-0.5",
    revision="6e397429f0ab91528ebac915004eba5e763fcdd4",
)
for item in inspection.input_contract.inputs or ():
    print(f"{item.name}: {item.shape} ({item.modality})")
```

## Deploy

Deploy the pinned public checkpoint to the robot registered with your organization:

```bash your computer theme={null}
servo model deploy hf://PerceptronAI/Isaac-0.5 --revision 6e397429f0ab91528ebac915004eba5e763fcdd4 --robot robot-1 --wait
```

## Import and deploy a fine-tune

Import the fine-tune at the same pinned revision you inspected, then deploy the returned model ID:

```bash your computer theme={null}
servo model import hf://ORG/REPO --revision COMMIT
servo model deploy mdl_YOUR_MODEL_ID --robot robot-1 --wait
```

The public checkpoint can be deployed directly from its `hf://` path as shown above. Preserve the
entire fine-tuned snapshot and import one pinned revision so its relative weight and processor
paths stay aligned.

## Send observations and receive actions

On the robot computer, construct `Session.predict(inputs=...)` with the exact keys, shapes, and
dtypes from the deployed policy. The [quickstart](/quickstart#choose-a-model) prints those resolved
inputs and shows the shared prediction flow.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.