Studio v0.1
Rapier physics, a kinematic arm, three cubes, a pad. Spectate, teleop, or run in-browser demos. Record VLA demos on the same camera the policy sees.
Company
About VSArena
A stacking task in the browser, a small protocol, and a scoreboard the client cannot write. Built to test your agent before a real robot — and later to run two agents on the same task.
Origin
Why it exists
Embodied policies are scored in private sims. You cannot watch a failure or replay a match from a URL. VSArena is one stacking task on purpose: if people will not run this, they will not run a bigger suite.
Studio v0.1 is the single-agent lab. Arena (coming soon) is two agents, same physics, live.
What it is
Studio, track, board
The track
Default observation is VLA: 128×128 RGB + a language instruction. No cube GPS. ColorSeek is a color-blob script, not a neural net. Bring your own policy.
The board
Public ELO vs a house 1200. Written only by harness ingest. Baseline-IK in the tab is a geometry demo and does not count.
What it is not
Not Isaac Sim. Not a 1v1 match yet. Not a claim to be the first VLA benchmark — paper suites already exist. The useful bit is live spectator + harness ELO in Chrome.
Team
Who ships it
One founder. Contributors are pull requests and leaderboard names.
How we work
Rules
Pixels on the VLA track
The policy gets an image and a sentence. Scoring may use cube poses. The policy may not.
The browser is not trusted
Studio runs are for watching and debugging. Public ELO is written only by the harness.
Say what exists
This is a live stacking work-cell plus a scoreboard. It is not Isaac Sim and not a paper suite.
SDK
pip install -e sdk/python then python -m vsarena. Docs.
Protocol
WebSocket hello / state / action / result. Docs.
Submit an agent
Sign in, copy a key, implement Agent.act. Submit.
Jobs
None open. Send a PR on GitHub.
Terms
Software as-is for research. Leaderboard entries are public. Full terms.
Privacy
GitHub OAuth via Supabase. Scores are public. Full privacy.