> ## 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.

# REKI

> Inspect and deploy Kyutai REKI-2B for Franka LIBERO observations and actions.

## Checkpoint

REKI's released LIBERO checkpoint is for a single Franka Panda arm. It uses the two 448×448
camera views and the native end-effector state layout below. The checkpoint weights are licensed
CC BY-NC-SA 4.0; the [REKI model card](https://huggingface.co/kyutai/reki-2b-libero) describes
the license and evaluation limitations.

```bash title="Your computer" sample=verified theme={null}
orchestral models inspect reki hf://kyutai/reki-2b-libero@589bd83a1237d45764fec65830a134ed8a5e3b69
```

```text title="Sample output" theme={null}
Inputs:
  camera_top: RGB uint8 (448, 448, 3)
  camera_wrist_left: RGB uint8 (448, 448, 3)
  state: float32 (8,)

Outputs:
  actions: float32 (10,7)
```

## Inputs and output

Use the camera keys listed below. Missing or extra keys raise an error.

| Direction | Name | Shape | Type |
| - | - | - | - |
| Input | `camera_top` | `(448, 448, 3)` | RGB `uint8` |
| Input | `camera_wrist_left` | `(448, 448, 3)` | RGB `uint8` |
| Input | `state` | `(8,)` | `float32` |
| Session context | task instruction | text | passed as `instruction=` when opening the session |
| Output | `actions` | `(10, 7)` | `float32` |

The state is `[eef_pos(3), eef_axis_angle(3), gripper_qpos(2)]`. Each prediction returns ten
delta end-effector actions. The model keeps recent history in register memory, so create one
Orchestral session per episode or pass `episode_start=True` when starting a new episode in an
existing session.

## Deploy and predict

Deploy the checkpoint URI directly. Orchestral resolves, imports and prepares it before bringing up the serving deployment.

```bash your computer theme={null}
orchestral deployments create hf://kyutai/reki-2b-libero@589bd83a1237d45764fec65830a134ed8a5e3b69 --robot robot-1 --wait
```

```text title="Sample output" theme={null}
jolly-tangerine-snail-rdik  ACTIVE · ready
ReKi 2B · robot-1
Timeline: requested 12s ago → starting worker 8s ago → ready now
```

```python robot computer theme={null}
import numpy as np
import orchestral as ol

client = ol.Orchestral()
deployment_id = input("Deployment ID from the CLI output: ").strip()
deployment = client.deployments.get(deployment_id)

inputs = {
    "camera_top": np.zeros((448, 448, 3), dtype=np.uint8),
    "camera_wrist_left": np.zeros((448, 448, 3), dtype=np.uint8),
    "state": np.zeros((8,), dtype=np.float32),
}

with deployment.session() as session:
    result = session.predict(inputs=inputs, instruction="put the mug on the plate")

print(result.actions.shape)  # (10, 7)
```

## Imported policies

Inspection reads REKI’s saved native training configuration. The server checks the loaded
policy’s camera, state, and action contract against that metadata and runs the upstream policy
preprocessing and output transforms.


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