DEXTEROUS AUTONOMY
[MASTERY LOOP] // A PATH TO MASTERY FOR DEXTERITY

MOST ROBOT STACKS STOP AT DEPLOY. WE START THERE — MEASURED IMPROVEMENT ON THE JOB UNTIL EVERY SKILL CLEARS THE MASTERY BAR: HUMAN COMPLETION, HUMAN THROUGHPUT, AND LOWER COST THAN HUMAN LABOR.

SHARE YOUR USE CASETHE MASTERY BAR >

SHOW // TRAIN // DEPLOY // MASTER

DWORK CELL / BIMANUAL — PARALLEL JAWLIVE █
TELEOPERATIONDEMONSTRATION DATATRAINING APISSTAGED ROLLOUTSUCCESS DETECTIONAUTO RESETSRESIDUAL RLINTERVENTIONS / HRCOMMAND CENTERTELEOPERATIONDEMONSTRATION DATATRAINING APISSTAGED ROLLOUTSUCCESS DETECTIONAUTO RESETSRESIDUAL RLINTERVENTIONS / HRCOMMAND CENTER
SKILL v12 ON DEPLOYMENT
[01] // THE BAR

ROBOTS MAKE ECONOMIC SENSE ONLY AFTER MASTERY.

A PATH TO MASTERY IS A SET OF KPIS. UNTIL ALL THREE HOLD, AUTOMATION IS A SCIENCE PROJECT — NOT A REPLACEMENT FOR PHYSICAL WORK.

  1. 01 // TASK COMPLETION≥ HUMAN

    THE ROBOT FINISHES THE TASK AT LEAST AS RELIABLY AS A PERSON — SUCCESS THAT HOLDS IN THE REAL CELL, NOT JUST IN A DEMO.

  2. 02 // TASK THROUGHPUT≥ HUMAN

    IT KEEPS PACE WITH THE LINE. MATCHING COMPLETION IS NOT ENOUGH IF THE ROBOT IS TOO SLOW TO REPLACE THE WORK.

  3. 03 // AMORTIZED COST≤ HUMAN

    ONCE PERFORMANCE AND THROUGHPUT CLEAR THE BAR, THE FULLY LOADED COST HAS TO COME IN AT OR BELOW HUMAN LABOR.

TRACKED ON EVERY SKILL, EVERY VERSION
  • SUCCESS RATE%
  • CYCLE TIMESEC / UNIT
  • INTERVENTIONSPER HOUR
  • COSTPER SUCCESSFUL UNIT

“DEPLOY AND FORGET” HAS A NUMBER: INTERVENTIONS PER HOUR TRENDING TO ZERO.

01 // HYBRID STACK

  • A BEHAVIOR TREE IS THE SCAFFOLD. CLASSICAL PLANNING HANDLES FREE-SPACE MOTION, APPROACH, RETREAT, AND RESETS.
  • LEARNED POLICIES TAKE ONLY THE CONTACT-RICH SEGMENT — WHERE DEXTERITY ACTUALLY LIVES.
  • RESIDUAL RL RIDES ON TOP OF THE BASE POLICY TO PUSH PAST HUMAN PERFORMANCE. HARDWARE-NEUTRAL BY DESIGN.

02 // MEASURED MASTERY

  • SHOWING, TRAINING, AND DEPLOYING ARE TABLE STAKES. WHAT HAPPENS AFTER DEPLOYMENT IS THE PRODUCT.
  • EVERY INTERVENTION BECOMES A CORRECTION. EVERY ROLLOUT BECOMES TRAINING SIGNAL.
  • EACH SKILL HAS A CURVE — FROM NOVICE TO MASTER — PLOTTED AGAINST THE MASTERY BAR.

  1. 01

    SHOW

    DEMONSTRATIONS SET A BASELINE FOR THE TASK.

    • >TELEOP — LEADER ARM, SPACEMOUSE, VR
    • >CAMERA CALIBRATION WIZARD
    • >AUTO EPISODE SEGMENTATION
    • >DEMO COVERAGE CHECKS
  2. 02

    TRAIN

    ONE-CLICK RECIPES TURN DEMOS INTO A BASE POLICY.

    • >ACT, DIFFUSION, π0.5 FINE-TUNE
    • >LEROBOT-FORMAT DATASETS
    • >OFFLINE EVAL ON HELD-OUT DEMOS
  3. 03

    DEPLOY

    THE POLICY GOES INTO THE REAL CELL — SAFELY, IN STAGES.

    • >SHADOW → SUPERVISED → AUTONOMOUS
    • >CANARY VERSIONS
    • >ONE-CLICK ROLLBACK
    • >SAFETY ENVELOPE
  4. 04

    MASTER

    MEASURED IMPROVEMENT ON THE JOB UNTIL THE SKILL CLEARS THE MASTERY BAR.

    • >SUCCESS DETECTION
    • >AUTOMATIC RESETS
    • >RESIDUAL RL, HUMAN-IN-THE-LOOP
    • >INTERVENTIONS CAPTURED AS CORRECTIONS
    • >KPI TRACKING

THE SKILL REPO

EACH SKILL IS A VERSIONED PROJECT: TASK SPEC, SUCCESS DETECTOR, RESET PROCEDURE, SAFETY ENVELOPE, DATASETS, POLICY LINEAGE, AND EVAL HISTORY AGAINST THE BAR. EVERY LOOP IS A COMMIT.
01 // SKILL

COMMAND CENTER

THE DESKTOP APP. DATA, INTERVENTIONS, DATA MIXTURES, AND THEIR IMPACT ON SUCCESS RATE, CYCLE TIME, AND COST — IN ONE PLACE.
02 // DESKTOP

CELL RUNTIME

AN EDGE AGENT ON THE ROBOT'S COMPUTER. OWNS REAL-TIME CONTROL, SAFETY LIMITS, THE WATCHDOG, AND LOGGING — AND KEEPS RUNNING WHEN THE LAPTOP CLOSES.
03 // EDGE

TRAINING BACKEND

YOUR LOCAL GPU OR OUR TRAINING APIS — FINE-TUNE FOR YOUR USE CASE WITHOUT OWNING THE CLUSTER.
04 // API

AN AGENT THAT ORCHESTRATES

IT RECORDS, TRAINS, EVALUATES, AND CLUSTERS FAILURES — THEN PROPOSES THE NEXT STEP. IT NEVER SENDS JOINT COMMANDS, AND ANYTHING THAT MOVES HARDWARE NEEDS HUMAN APPROVAL.
05 // AGENT

THE HARD PLUMBING

REAL-WORLD LEARNING FAILS ON PLUMBING, NOT ALGORITHMS. AUTOMATIC RESETS, RELIABLE SUCCESS DETECTION, SAFE EXPLORATION BOUNDS, AND REGRESSION CHECKS ON EVERY NEW VERSION.
06 // RELIABILITY

START WITH A MASTERY ENGAGEMENT: ONE TASK, ONE CELL. TELL US THE PHYSICAL TASK AND WE'LL RUN THE LOOP UNTIL THE ROBOT MATCHES HUMAN COMPLETION AND THROUGHPUT — AT AN AMORTIZED COST AT OR BELOW HUMAN LABOR — WITH THE MASTERY CURVE TO PROVE IT.