microdex.ai

Execute your own cube moves.

Open Cube Lab to build a turn sequence, solve a scramble, and check the actual cube state.

Open Cube Lab →
PRODUCT DIRECTIONS / OPEN EXPLORATION

One task is a beginning.

A proposed path from the Frontier replay lab to repeatable robotics experiments, shared tools, and new ways to move.

WHERE WE ARECONCEPT ROADMAP

Build on something you can inspect.

Today, MICRODEX offers one recorded cube demo, editable playback profiles, and browser-local results. The ideas below describe possible next products. They are not available features or promised launch dates.

TODAY

Explore a replay

Run Frontier, create a JSON profile, and compare local playback results.

NEXT TO VALIDATE

Execute a real task

Introduce bounded robot actions, repeatable starting conditions, and meaningful task scores.

FUTURE DIRECTIONS

Expand the environments

Carry the same experiment workflow into manipulation, leg control, and whole-body motion.

PRIORITY 01 / THE FOUNDATION

Make the task executable.

01 / POLICY ARENA

A fair starting line.

Give every controller the same task, initial state, action limits, and run budget. Compare success, completion time, and invalid actions across repeatable trials.

First candidate: a small, deterministic manipulation task with a recorded observation/action log for every attempt.

02 / CURRICULUM CHALLENGES

Build up one skill at a time.

Progress from a single controlled turn to a sequence of object movements. Use clear pass conditions and replay failures to see where a policy needs work.

Useful for testing whether an improvement survives harder starting conditions.

03 / GRIPPER PICK & PLACE

Put the right object in the right place.

Pick up simple objects, move them between targets, and release them reliably. Evaluate placement accuracy, dropped objects, and collisions.

A practical next task for comparing perception and control choices.

PRIORITY 02 / THE EXPERIMENT WORKFLOW

Make every run useful.

04 / IMITATION DATASET REPLAY

Inspect the demonstration.

Import compatible demonstration datasets, align observations with actions, and compare successful and failed segments before training a policy elsewhere.

A future extension beyond the single recording currently bundled with Frontier.

05 / DEVELOPER SDK

Bring a controller to a documented interface.

Provide an observation/action contract, task fixtures, and a command-line evaluation runner. Export reproducible results with the environment and policy versions attached.

Requires an executable task environment and a bounded runtime first.

06 / PRIVATE TEAM BENCHMARKS

Compare a change with your own baseline.

Let teams maintain private task sets, review policy revisions, and spot regressions through shared run records and controlled evaluation settings.

Would require accounts, storage, permissions, and independently evaluated runs.

07 / CLASSROOM LABS

Turn a run into a lesson.

Create guided exercises in sensing, planning, and control. Students change one setting, record a prediction, run the task, and explain the result.

Start with downloadable lesson material and local experiments; add classroom management only when needed.

PRIORITY 03 / NEW ENVIRONMENTS

Give the experiment room to move.

08 / SIM-TO-REAL EXPORT

Prepare an artifact for a real robot.

Export joint mappings, units, timing, and bounded commands for a supported hardware adapter. Include a dry-run report so a developer can inspect compatibility before any physical execution.

Future export workflow. Hardware control and validated transfer from simulation are not available.

09 / LOCOMOTION & BALANCE

Find a stable next step.

Explore standing balance, controlled stepping, and recovery from small disturbances in the proposed Stride leg environment.

Needs a contact-aware dynamics model, repeatable tests, and new evaluation measures.

Meet Stride / Leg
10 / WHOLE-BODY MANIPULATION

Reach while the rest of the body helps.

Coordinate posture, balance, and arm movement in the proposed Atlas body environment. Begin with a stationary reach, then investigate carrying and placing.

A later concept that depends on reliable manipulation and locomotion foundations.

Meet Atlas / Body
THE NEXT USEFUL STEP

Try the workflow that exists today.

Build a profile, run Frontier, and inspect the local result. That loop is the starting point for deciding which future task is worth building.