[HTML version](https://blokk.cloud/competition)

# CubeStack

Run code on robots for free before beta access. Introducing.

## 1. Learn the rules

Train a policy to stack as many cubes as possible in our pods.

## 2. Submit

Log in and paste your link. Or have your agent do it.

## 3. Queue

Your run is queued onto the robot. See where you are in line.

## 4. Watch

Watch what's happening in the pod, any time.

## 5. Climb

Get a full eval report, iterate and climb the leaderboard.

## Stack more, unlock more

Community rewards open for everyone when anyone stacks them. Yours open when you do.

### Community

- 1 cube: Free pod access. Free runs now, on a Jetson Orin Nano.
- 5 cubes: GPU access. A free GPU connected to the arm. Queue bigger, more complex policies.
- 8 cubes: Season 2. New rules. New objects in the pod.

### Individual

- 1 cube: Certificate. Add it to LinkedIn, your CV or your site.
- 5 cubes: Skip the beta queue. Straight to the front of the waitlist.
- 8 cubes: Free pod hours. Time on real robots, on us.

## Onboard your agent

The page gives you this prompt to paste into your coding agent:

```text
Set me up for the Blokk CubeStack competition. Read https://blokk.cloud/competition.md for the rules and task spec. Install MuJoCo 3.3.7 on Python 3.12 and get the SO-101 model from MuJoCo Menagerie. Blokk's own scene isn't published yet, so build a stacking scene from the spec: eight 25 mm cubes placed at random in the stated play area. Train a policy that takes the 18 state values and the 1280×720 wrist camera image, returns 6 absolute joint targets, runs on CPU with no network, and stacks as many cubes as possible in 120 seconds. Evaluate it in sim, publish it as a public Hugging Face model repo on my account, and give me the link to paste into the queue.
```

## CubeStack rules

Draft, task spec included. A full rulebook is coming. Until then, these keep things fair for everyone. Full text: [rules](https://blokk.cloud/rules.md).

### Task

A real SO-101 arm in an enclosed cell. Eight orange 25 mm cubes are placed at random, without overlapping.

- Cell interior: 729.4 × 582.4 × 555 mm.
- Play area, inside the 40 mm perimeter barriers: 649.4 × 502.4 mm.

### Fair use

1. One run queued per account at any time.
2. Each run lasts two minutes. Then the arm returns home and the cell resets.
3. One account per person or team. No extra accounts to jump the queue.
4. Submit policies as public Hugging Face model repos. They may only use the observations we provide.
5. No policies designed to damage the arm, the cell or the cameras. We stop any run that looks unsafe.
6. Scores are final once the evaluation report is published. Appeals within 7 days.
7. We may record runs and publish recordings and scores on the leaderboard.
8. We may disqualify any run or account that breaks these rules.

### Scoring

Your score is the height of the tallest stable stack at the end of the run, in cubes. Ties go to whoever placed their final cube first. An overhead depth camera measures it automatically once the arm is at rest, counting stacks up to 6 cubes.

### Policy interface

Your policy exposes `predict(observation)` and returns the action. It is called once per new wrist camera frame (30 fps). There is no fixed control rate.

- `state`: 18 values. 6 joint positions, 6 previous joint targets, 6 joint velocities.
- `image`: the wrist camera, 1280 × 720 RGB, uncropped. The only camera the policy sees.
- Action: 6 absolute joint position targets.

Joint order: `shoulder_pan`, `shoulder_lift`, `elbow_flex`, `wrist_flex`, `wrist_roll`, `gripper`. Joints 1–5 are in degrees, the gripper in % of its range (0–100). Velocities are in degrees per second (gripper: percentage points per second). Actions use the same order and units.

### Runtime

Policies run on the robot's Jetson in an isolated sandbox: CPU only, no network access. Bundle all dependencies with your submission.

### Submission

A link to a public Hugging Face model repo, like `https://huggingface.co/owner/model`. One active submission per account. The package format (required files and manifest) is still being finalised.

### Simulation

MuJoCo 3.3.7, Python 3.12. The arm model is the SO-101 from MuJoCo Menagerie ([robotstudio_so101](https://github.com/google-deepmind/mujoco_menagerie/tree/main/robotstudio_so101), Apache-2.0). Our CubeStack scene and task environment are not yet published.

## Links

- [See leaderboard](https://blokk.cloud/leaderboard)
- [Full rules](https://blokk.cloud/rules)
- [Open console](https://blokk.cloud/console)
