Teach AI how the physical world gets fixed.
ZOL transforms repair videos, technical manuals, diagrams, and technician knowledge into grounded procedures that can guide people today - and power embodied AI tomorrow.
ZOL procedural engine
2020 Toyota Camry - engine air-filter replacement
Pipeline
Raw input
Toyota Camry repair video
Perception
Parts and tools detected
Extraction
Steps and state changes extracted
Structure
Procedural action graph created
Inference
Unseen video analyzed
Guidance
Current step: Release housing clips
Live procedure state
Front clip
ReleasedRear clip
Not visibleHousing cover
ClosedProcedure status
2 of 10 steps confirmed
Current step
Release housing clips
Next action
Lift the housing cover only after every retaining clip is confirmed released.
Most AI can describe an image. ZOL understands a procedure - what happened, what changed, what must happen next, and when the system should stop and ask for more evidence.
The world's most valuable physical knowledge is trapped in unstructured content.
Millions of repair demonstrations exist across videos, manuals, diagrams, forums, and technician experience.
Videos show how people perform the work, but they are difficult for machines to search, verify, or execute.
Manuals contain the official procedure, but they do not show the visual variations technicians encounter in real environments.
Existing AI can answer questions about a frame, but it rarely maintains procedural state across an entire physical task.
Robots and AI agents need more than instructions - they need grounded actions, dependencies, state transitions, failure conditions, and uncertainty.
The data already exists. The missing layer is the system that converts it into machine-usable procedural intelligence.
From demonstrations to executable knowledge.
Visual understanding
Identifies tools, parts, hand-object interactions, actions, and important visual evidence across time.
State-change tracking
Understands that a component changed from attached to removed, closed to open, empty to installed, or unsecured to secured.
Procedural action graphs
Converts demonstrations into structured steps with preconditions, resulting states, valid transitions, common mistakes, and completion conditions.
Grounded repair memory
Connects visual demonstrations with manuals, diagrams, product information, specifications, and previously verified examples.
Real-time guidance
Uses the current visual state and completed-step history to explain what the technician should do next - or request a clearer view when evidence is insufficient.
One intelligence layer across the entire procedure.
Step 1
Learn
ZOL ingests repair videos, manuals, diagrams, and parts information.
Step 2
Structure
It extracts actions, objects, timestamps, physical state changes, dependencies, and common variations.
Step 3
Verify
It compares demonstrations against authoritative technical documentation and flags disagreements or missing evidence.
Step 4
Reason
During a new repair, ZOL determines the current step, retrieves similar demonstrations, and validates the action against the procedure graph.
Step 5
Guide
It provides the next instruction, highlights risks, and explains what is visually confirmed versus uncertain.
Watch ZOL understand a repair it has never seen before.
2020 Toyota Camry - engine air-filter replacement, analyzed from a held-out silent repair video.
Held-out silent repair video
No audio. No annotations. No prior exposure.
Detected
Technician is releasing the front housing clip
Physical state
Front clip
ReleasedRear clip
Not visibleHousing cover
ClosedProcedure status
2 of 10 steps confirmed
Steps advance only when the required evidence is visually confirmed.
Next action
Show the rear side of the housing and confirm that the second clip is released.
Reason
The housing should not be lifted until all retaining clips are released.
Evidence
- 2 similar demonstrations
- 1 relevant manual section
- 89% visual confidence
A chatbot answers. ZOL maintains procedural state.
General visual chatbot
ZOL
Built for actions, not captions.
The language model interprets visual evidence and communicates naturally. ZOL's procedural engine controls task state, retrieves the correct domain knowledge, validates step order, and prevents unsupported actions from advancing the procedure.
Multimodal extraction
+
Procedural memory
+
State-transition graph
+
Vehicle-specific retrieval
+
Deterministic safety validation
+
Frontier model reasoning
Every procedure makes the system harder to replicate.
Every processed video adds new camera angles, technician styles, part appearances, and real-world variations.
Every verified procedure adds structured preconditions, state changes, errors, and completion evidence.
Every guided session produces examples of where models succeed, hesitate, or make incorrect predictions.
Every vehicle and repair expands a reusable procedural ontology.
The resulting dataset can support guidance, model training, benchmarking, and robotic planning.
We are not collecting more repair content. We are converting repair content into structured physical intelligence.
Guide technicians today. Train physical AI tomorrow.
Today
- Technician repair copilot
- Visual step verification
- Next-action guidance
- Training and onboarding
- Remote expert assistance
- Quality-control documentation
Tomorrow
- Embodied-AI training data
- Robot task planning
- Simulation and evaluation
- Procedural benchmark creation
- Cross-vehicle state representations
- Human-to-robot skill transfer
Physical AI needs evals, not just demos.
Every guided session is scored against the questions that determine whether a system truly understood the work.
Did the system identify the correct action?
Did it recognize the affected tool and component?
Did it detect the physical state change?
Did it understand the required earlier steps?
Did it predict a valid next action?
Did it catch a skipped or incorrect step?
Did it recognize when visual evidence was insufficient?
Did the procedure finish in a correct and safe state?
See ZOL understand a repair.
Tell us what you work on and we will walk you through the system on a real procedure.
The next generation of AI will not just understand language. It will understand work.
ZOL is building the intelligence layer that turns human demonstrations into structured, verifiable, and actionable physical knowledge.