Skip to content

Platform

Three ways to train on maintenance decisions.

Practice a decision in an environment, learn from an expert trajectory, or evaluate a model against a defined task.

01 / Environment specification

RL Environments

An episode begins with a work order and a limited view of the machine. The underlying fault stays hidden. Inspection requests reveal new evidence, while repair actions change the machine state and consume maintenance time.

Inputs
Equipment configuration, fault state, permitted actions, safety prerequisites.
Outputs
An observation after each action and a terminal record of the maintenance decision.
Illustrative record / JSON
{
  "observation": "drive_end_vibration_above_baseline",
  "action": "request_envelope_spectrum",
  "next_observation": "outer_race_frequency_harmonics"
}
02 / Annotated trajectory

Expert Demonstrations

A demonstration preserves the order in which evidence became available. Technicians describe why an observation matters, which alternatives remain plausible, and what would justify interrupting production.

Inputs
A scoped scenario, an equipment-qualified contributor, and a structured capture protocol.
Outputs
An ordered trajectory of observations, decisions, and expert rationales for supervised or preference training.
Illustrative record / JSON
{
  "observation": "localized_drive_end_impacts",
  "hypothesis": "bearing_outer_race_damage",
  "decision": "schedule_isolated_inspection",
  "rationale": "Confirm physical condition before committing to replacement."
}
03 / Task and grading rubric

Evaluations

A grading specification ties each criterion to observable behavior in the trajectory. Reviewers can distinguish a well-supported decision from a lucky diagnosis and recognize alternative sequences that respect the same constraints.

Inputs
A held-out scenario, agent trajectory, and reviewed grading criteria.
Outputs
Criterion-level findings with evidence from the agent’s decisions.
Illustrative record / JSON
{
  "criterion": "repair_justification",
  "evidence_required": [
    "localized_findings",
    "confirmation_plan"
  ],
  "review": "Did the requested intervention follow from the observed evidence?"
}

Scoping a pilot

Start with the decision your model needs to make. Kosho can discuss the machine family, fault coverage, observation format, intended training use, and the evidence you need to evaluate a pilot.

Research inquiries

Build better physical-world intelligence.

Explore expert-informed maintenance environments and evaluation data for your model-training needs.