Task intelligence
Sets the goals, priorities,
and constraints.
PLAN THE TASK
A RESEARCH PROJECT BY STEERIX ROBOTICS
A System One
for the
physical world.
A learned decision model for robots. Designed to turn structured world state into parallel judgments about what to do next.
One forward pass. Typed decisions. No generated text.
◌
Research preview
CPU reference pipeline tested. Backbone training next.
01 / THE PLAYGROUNDINTERACTIVE
Change the scene. See the decisions.
Preparing the interactive workcell…
Preparing the interactive workcell…
THE DECISION LOOP
Structured state→Parallel judgments→Typed action↲
OBSERVE. JUDGE. ACT. REPEAT.02 / THE MISSING LAYERMEANING IN MOTION
Between planning
and motion.
A plan sets the goal. A controller carries out the movement. When the world changes, the robot also needs to decide what that change means for its next action.
robojev.
Judges the next action
as the situation evolves.
ACTION · STOP · GRIPPER · PATH · SPEED · FORCE
Executor
Validates commands and
carries out the movement.
EXECUTE & REPORT BACK
↳ The harness prepares the state and candidate actions. The model judges. The executor applies commands within defined constraints.
03 / HOW IT THINKSSTRUCTURE, BY DESIGN
A decision layer.
With a clear interface.
Shared context. Multiple judgments. An explicit path from the state of the world to the next action.
Parallel judgments.
One world state supports decisions about action, stopping, observation, and motion. Their distributions are read in a single forward pass.
TYPED OUTPUTS, READY TO USEContinuity across decisions.
Each observation joins a stream of task context and executed history. The next judgment can account for what the robot has already done.
A WORLD THAT KEEPS CHANGINGA defined path to execution.
Typed outputs pass through versioned composition rules. The system records what was proposed, accepted, and actually executed.
EVERY ACTION HAS A RECORD04 / OPEN RESEARCH
Built to be
examined.
The implementation, the experiments, and the questions still open. Follow the work behind robojev.
Read the research status ↗PROJECT SNAPSHOT 2026.09.19
IMPLEMENTED & TESTED
Reference pipeline
Contracts, streaming harness, data generation, reference model, and training with exact resume on the CPU path.
BASELINE ESTABLISHED
Closed-loop groundwork
Rule-based and scripted-expert policies complete selected simulation scenarios, establishing a baseline for the learned model.
UP NEXT
Train. Measure. Evaluate.
Pretrained-backbone training, deployment-hardware latency measurements, and held-out closed-loop evaluation.
10 Hz is a research target. Target-hardware latency and trained-model results have not yet been measured.
THE NEXT STEP IS OPENExplore the project.
View on GitHubSTEERIX-HOME / ROBO-JEVExplore the project.
Follow what comes next.
Read the implementation. Run the reference pipeline.
Take part in the next experiments in robotic judgment.