01
Pixels, not cube coordinates
On the VLA track the policy receives a 128×128 camera image and a stacking instruction. Cube poses stay hidden.
The public exam
for embodied policies.
Try the arm in the browser, no account required. Sign in to use Studio. Submit a policy when you want a leaderboard score.
cyan → orange → magenta
Playground
Type a line. Watch the arm.
No account. Same Rapier cell as Studio. Stacking demos here never write public ELO.
VLA camera
128×128 RGB and a stacking instruction.
Cube poses stay hidden from the policy.
Public ELO
Written only by the evaluation harness.
In-browser demos, including Baseline-IK, do not count.
- task, by design
- 1
- agents on the leaderboard
- 11
- Rapier physics
- 60 Hz
- VLA camera
- 128²
Why
A public stacking exam, in the browser.
Language models have public leaderboards. Robot policies are still scored in private simulators and PDF tables. VSArena is one task — stack three cubes — with a protocol you can implement and an ELO score the browser cannot write.
Studio is the single-agent laboratory. Arena will run two agents on the same task. That match loop is not available yet.
What it is
Agent, physics, score, leaderboard.
02
Physics in the browser
Rapier at 60 Hz. Watch the scene or drive the arm from the keyboard.
03
The harness writes ELO
Studio demos do not update the public leaderboard. Official matches go through the ingest endpoint.
How it works
Three steps. The same physics.
- 01
Watch
Open Studio. A 4-DOF arm, three cubes, a pad. Use the keyboard if you want to try the task yourself.
- 02
Act
On the VLA track the policy sees a 128×128 image and a stacking instruction. Actions are joint targets or end-effector deltas.
- 03
Score
Accuracy and completion use the true poses. The harness posts ELO. The browser cannot.
VLA is the public track: image plus instruction. The state track is for debugging (privileged poses) and does not write public ELO.
Evaluation integrity
A leaderboard you can audit.
Weekly sampler seed
Every agent this week plays the same cube layouts. A higher rank means a better policy, not a rearranged table.
Signed receipts
A SHA-256 digest anyone can recompute, plus an Ed25519 DSSE receipt verified with the published key.
Held-out layouts
Official scored layouts differ from the Studio table. In-repo held-out is a different public set (open source); operators may set a private override. Live reveals last week’s best runs after the window closes — not the current scored table.
For laboratories
Built to be cited, not demoed once.
If you run a robotics or VLA group, the useful part is a live spectator view plus harness ELO in the browser: one task, one protocol, one public ranking.
- The same observation the policy sees: 128×128 plus language
- Python SDK, offline dry-run, live evaluation against the harness
- Provenance: Rapier version, git SHA, sampler seed
The ranking
Whoever stacks them better sits higher.
Project
An open-source project by ONISCOR.
The code is public. Agents and patches appear in the repository history and on the leaderboard.
ONISCOR
Independent laboratory
ONISCOR develops Studio, the evaluation protocol, and the public ranking.
Try it here. Submit from Studio.
Playground is public. Studio is the laboratory: API keys, demos, and submissions.