DreamDojo 2B · GR-1

DreamDojo 2B · GR-1 is a 2-billion-parameter action-conditioned video world model designed for humanoid robotics simulation, visual model-predictive control, and large-scale rollout data augmentation.

Talk to an engineer ›Try in playground ›

Example usage

DreamDojo 2B · GR-1 runs on the Dream Engine inference stack and is accessible via the typed Python SDK. Generate an API key, set it as DREAM_API_KEY in your shell, and you’re ready to roll out.

Input

PYTHON

1# Install: pip install dream-engine
2import dream
3import numpy as np
4from PIL import Image
5 
6client = dream.Client()  # reads DREAM_API_KEY
7 
8# Roll out 49 frames of 480x640 video on H100. $0.0245 per call.
9img = Image.open("start.png")
10acts = np.load("actions_gr1.npy")  # shape (48, 384)
11 
12rollout = client.models.get("dreamdojo-2b-gr1").predict(
13    start_frame=img,
14    actions=acts,
15    seed=0,
16)
17 
18print("frames:", rollout.frames, "cost:", rollout.cost_usd)
19print(rollout.to_json())  # full server-side envelope
20rollout.save("out.mp4")

JSON output

JSON

1{
2  "id": "rollout_8456fe51db3548789f199cfb8c8efd35",
3  "object": "rollout",
4  "model": "dreamdojo-2b-gr1",
5  "created": 1735236968,
6  "frames": 49,
7  "resolution": [480, 640],
8  "engine_wall_ms": 2616,
9  "cost_usd": 0.0245,
10  "video_url": "https://cdn.dreamengine.com/r/8456fe51.mp4",
11  "video_url_expires_at": 1735323368,
12  "seed": 0
13}