digiarena (PyPI: digiarena)

Python SDK for DigiArena — browse hosted robot-simulation scenes, deploy them as cloud gyms, and submit to benchmarks.

pip install digiarena
# Camera images are rendered on your machine; add this for image observations:
pip install 'digiarena[render]' && python -m playwright install chromium

import digiarena as da

da.login()                                   # paste an API key from digiarena.org → Settings
# Continue the exact browser attempt without resetting:
with da.attach("hosted-session-id") as env:
    obs, info = env.observe()
    print(info["episode_id"], obs["rgb"].keys())

# Or launch an independent session from an accessible scene in one call
# (env.close() also tears the session down):
with da.make("alice/kitchen_001", task="pick_mug_to_sink") as env:
    obs, info = env.reset(seed=0)
    obs, reward, terminated, truncated, info = env.step(env.action_space.sample())

# Or keep the handle for the viewer URL, logs and lifetime extension:
scenes = da.scenes.list(query="kitchen")
gym = da.deploy(scene="kitchen_001", task="pick_mug_to_sink")
env = gym.connect()                          # → world_engine_client.RemoteEnv
obs, info = env.reset(seed=0)
obs, r, term, trunc, info = env.step(action)
gym.close()

CLI with the same surface:

digiarena login
digiarena scenes list --query kitchen
digiarena deploy --scene kitchen_001
digiarena leaderboard tabletop-v1
digiarena usage

Layering

digiarena depends on and reuses world-engine-client (connect(token, url) → RemoteEnv). This package covers everything from your account to that token — scene gallery, cloud deploy, benchmark submission, leaderboards, usage — and never reimplements the wire protocol.

• Light dependencies: numpy, Pillow (camera images) and Gymnasium (spaces), all via world-engine-client. No torch or opencv. import digiarena loads only the stdlib and numpy; Pillow and Gymnasium load on first use.
• Every call hits the public REST API (/api/v1/*) with Authorization: Bearer da_live_... — the SDK is a thin wrapper you could replace with curl.
• Typed errors with actionable "Possible fixes" (the World Engine style).
• Endpoints still in later platform milestones (benchmark submission) raise NotImplementedYetError with the server's own guidance — the SDK surface is stable ahead of them.

Scene content never transits your machine: there is no upload()/download(). Content is authored in the platform's browser editor; public scenes are used directly via da.deploy() without forking. Private scenes require ownership or platform admin access. Pass images=False to deploy/attach for state-only operation.

Camera images are drawn on your computer, not on the platform: the hosted session simulates physics and sends the poses with each step, and a headless Chrome on your machine renders the same scene the browser shows. This needs the render extra and a Chrome with hardware WebGL (chrome://gpu); without the extra, asking for images fails before any request. Scene assets are downloaded once per session into a temporary cache that env.close() deletes.

Remote simulation

da.attach(session_id) uses resume semantics and pauses the current attempt. env.observe() and env.render() never advance physics. env.reset(seed=...) starts a new episode; env.step(action) advances one negotiated control interval. env.close() releases control; gym.close() tears down a launched session. A live Python controller survives browser closure. Lost/uncertain actions are never retried automatically: reconnect or explicitly reset to recover.

Images are RGB uint8 arrays in obs['rgb'], keyed by configured camera name. Use cameras=['wrist'] for a subset and env.as_gymnasium() for the optional Gymnasium adapter.

Runnable examples are in the repository's examples/ directory (not part of the installed package): attach.py, reachable_pose.py, and deploy_policy.py. measure_remote.py records throughput and soak measurements. vla_smolvla.py runs a real VLA (SmolVLA via LeRobot, pip install 'lerobot[smolvla]') on your machine against a remote session, and measure_link.py reports what your network link does to stepping: steps per second, latency and bandwidth used, for state only, one camera and all cameras.

To check the whole path end to end, start Simulate in the browser editor, copy the session id from the Inspector's Python API card, and send the end effector to a pose while watching the browser:

python examples/reachable_pose.py SESSION_ID                       # small offset from the start pose
python examples/reachable_pose.py SESSION_ID --pos 0.45 0.0 0.30   # absolute position, metres
python examples/reachable_pose.py SESSION_ID --pos 0.45 0 0.3 --rot 3.14 0 0 --state-only
digiarena attach SESSION_ID --state-only                           # inspect without stepping

Development

pip install -e '.[img,gym]' pytest && python -m pytest

The tests use a stub control plane and a fake simulator; no platform is needed.

Credentials

DIGIARENA_API_KEY env var (CI/batch) → ~/.config/digiarena/credentials (written by da.login() / digiarena login, mode 0600). DIGIARENA_API_URL overrides the endpoint for staging deployments.