New GR00T N1.5 and π0.5 are live

One API for any VLA or world model

Send an observation from your robot or simulator, get an action chunk back. 10 models, one request schema, no per-model integrations.

Free to start · No credit card · Keys in 30 seconds

preview
POST /v1/act
sinkcup 0.97RECscene_cam · 224×224 · 30 fpsf 0000approach
"model": "physical-intelligence/pi0",
"instruction": "pick up the green cup and place it in the sink",
"observation": { "image", "state" }
200 OK · actions73 ms
txyzrollpitchyawgrip
0+0.000+0.035+0.044+0.051-0.053-0.1171.0
1+0.015+0.043+0.038+0.011-0.087-0.1191.0
2+0.029+0.045+0.027-0.031-0.110-0.1061.0
3+0.039+0.042+0.013-0.069-0.120-0.0801.0
4+0.044+0.034-0.003-0.098-0.115-0.0451.0
5+0.044+0.021-0.018-0.116-0.096-0.0041.0
6+0.039+0.006-0.031-0.120-0.066+0.0371.0
7+0.029-0.009-0.040-0.109-0.028+0.0741.0
8+0.015-0.024-0.045-0.085+0.014+0.1021.0
9-0.000-0.035-0.044-0.050+0.054+0.1180.0
streaming…

Built for the robots and simulators you already use

Franka PandaUR5eALOHAUnitree G1SO-100WidowX 250Kinova Gen3MuJoCoIsaac SimManiSkillGenesisLeRobot
Why OpenRobotics

Every robot model is its own integration project

Today

  • A different SDK, checkpoint format and runtime for every model
  • Hand-written action-space conversion and normalisation per robot
  • Your own GPUs to host each model, sized for peak load
  • Weeks of glue code before you can compare two policies

With OpenRobotics

  • One endpoint and one request schema for every VLA and world model
  • Your robot’s action space mapped for you, per embodiment
  • Hosted inference, billed per call, with spend limits per key
  • Swap π0 for GR00T by changing one string
How it works

Observation in, actions out

1 · InputRobot or simcamera frames, joint state and a language instruction
2 · OpenRoboticsOne APIauth, routing, action-space mapping, billing
3 · OutputAny VLA or world modelaction chunks or predicted rollouts
Quickstart

Your first call in under a minute

  1. 1

    Create an account

    Email and password. Nothing else.

  2. 2

    Create an API key

    Name it, optionally cap its spend, copy it once.

  3. 3

    Run the command

    The snippet on the right works today: it checks your key and returns its remaining credit.

Create a key
curl https://openrobotics.ai/api/v1/key \
  -H "Authorization: Bearer $OPENROBOTICS_API_KEY"
Runs today against this server · GET /api/v1/key
import os, base64, requests

obs = {
  "image": base64.b64encode(open("wrist_cam.jpg", "rb").read()).decode(),
  "state": [0.12, -0.48, 0.33, 0.0, 1.57, 0.0, 0.04],   # joints + gripper
}

res = requests.post(
  "https://openrobotics.ai/api/v1/act",
  headers={"Authorization": f"Bearer {os.environ['OPENROBOTICS_API_KEY']}"},
  json={
    "model": "physical-intelligence/pi0",
    "instruction": "pick up the red cup and place it in the sink",
    "observation": obs,
    "action_horizon": 50,
  },
)
actions = res.json()["actions"]   # 50 x 7 action chunk
Preview · POST /api/v1/act launches soon

Give your robot a brain in one request.

Create a key and call any VLA or world model through the same endpoint.

Get an API key