Task
Sets the goal and the order of work: the tables to serve, the tubes to prepare, the order to deliver.
FIXED FOR EVERY MODEL
A benchmark for real-time physical judgment.
Same robots, same situations, same controllers. Only the decision model changes. Pick a model and watch it decide, live in your browser.
robojev bench is the benchmark; Jev, d1, Decision-1, GPT-6 Luna, Solar Flash and robo-jev are models you can run in it.
01
Environments
Rule vs System 1 Move a block or tip a bottle mid-run, and watch a fixed rule and System 1 meet the same surprise.TESTSFollowing a part that moves or tips over mid-task.
System 1 drives Hand the truck to System 1 on a licence-exam course, or take the wheel yourself.TESTSLane and docking calls on an exam course.
Chooses where to be A G1 carries cartons across a forklift’s lane; watch System 1 decide when to cross and when to wait.TESTSWhether to cross a forklift’s lane, or wait.
Three a side Kick off a three-minute match; one call a second moves every duck on both teams.TESTSEvery move of six robots, once a second.
A waiter that reads the room Click a customer to have them order a drink, wave or call out, and see whom the waiter serves.TESTSNoticing who is calling, and what they want.
Runs the tube cell Watch System 1 scan, weigh, spin, uncap, aliquot and recap a day of sample tubes — and decide what to set aside.TESTSSpotting an odd sample and setting it aside.
Delivers through the lunch crowd A sidewalk robot delivers on real Gangnam streets; System 1 picks every manoeuvre, pass and crossing, and calls the operator only when it must.TESTSA way through the crowd, and when to call the operator.
02
How it works
Only the decision model changes.
Every environment splits the robot’s work into three layers. The task and the executor stay fixed; the decision model between them is the one you choose.
Decision model
Reads the robot’s state as it moves and answers typed questions: what next, who first, wait or go.
THE MODELS BELOW · YOUR MODEL
Executor
Carries each answer out with the environment’s own controllers.
FIXED FOR EVERY MODEL
- 01
Same harness
Every model gets the same state, the same questions and the same executor.
- 02
Real time
Robots ask as they move, once a second or faster. A late answer is carried out late.
- 03
Typed questions
Choice, score or yes/no, through OpenRouter’s Decisions API.
- 04
The robot’s senses only
No model sees the simulator’s ground truth. Where a stand-in replaces a perception model, the environment’s page names it.
03
Models
Choose one on any environment’s page.
The models the site runs answer through OpenRouter’s Decisions API. No scores yet: watch them decide.
Jev
by TypeSafeAVAILABLEA decision model served through OpenRouter’s Decisions API. It reads a description of the state and answers typed questions with probabilities.
Runs in all seven environments.
d1
by Liquid AIAVAILABLEA decision model served through OpenRouter’s Decisions API. It reads a description of the state and answers typed questions with probabilities.
Runs in all seven environments.
Decision-1
by MicrosoftAVAILABLEA decision model served through OpenRouter’s Decisions API. It reads a description of the state and answers typed questions with probabilities.
Runs in all seven environments.
GPT-6 Luna
by OpenAIAVAILABLEA decision model served through OpenRouter’s Decisions API. It reads a description of the state and answers typed questions with probabilities.
Runs in all seven environments.
Solar Flash
by UpstageAVAILABLEA decision model served through OpenRouter’s Decisions API. It reads a description of the state and answers typed questions with probabilities.
Runs in all seven environments.
robo-jev
by Steerix RoboticsON REQUESTSteerix Robotics’ own decision model, made for physical judgment, and open source. Steerix serves it directly rather than through OpenRouter, so its calls take another network path than the other models’, and turns it on for bench runs; while it is off, it cannot be chosen.
It runs in all seven environments.
Steerix Robotics also runs this bench. robo-jev runs on the same harness as every other model.
Bring a model
Any model that answers OpenRouter’s Decisions API can run in every environment.
- REQUIREMENT
- Takes a description of the state and choice, score and yes/no questions, and answers with probabilities.
- SAME HARNESS
- Every model runs on the same harness: state, questions and executor. The harness was first built with Jev, and each version of it is fixed.
- REAL TIME
- Robots ask once a second or faster, so answers should come back within about a second. Until an answer comes, the robot carries on with its last decision, as a real robot would.
- CONTACT
- admin@steerix.ai