Industrial robot arm lifting a crate in a warehouse aisle while a worker in a high visibility vest walks ahead

Train robots on episodes, not isolated frames.

YPAI turns human and robot demonstrations into synchronized, training-ready episodes with failures, interventions and recoveries intact.

Representative episode
Grasp Failure detected Operator recovery Accepted episode

A box around an object can train perception. It cannot teach the task.

Physical AI systems need more than isolated labels. They need to know:

A labelled frame

  • object class
  • object location

A complete episode

  • what the system observed
  • what instruction or objective applied
  • which action followed
  • what changed in the environment
  • whether contact occurred
  • when each task stage began and ended
  • whether the attempt succeeded
  • why it failed
  • whether a human intervened
  • how the system recovered

YPAI structures these relationships as complete episodes rather than disconnected frames.

The same holds for industrial vision: a visual anomaly becomes useful training data only when it is connected to the product, station, process stage, tolerance, viewpoint, inspection method and disposition.

The project is built around the decision the system must make.

That decision may be:

  • which object to manipulate
  • how to grasp or move it
  • whether an assembly is correct
  • whether a surface contains a defect
  • whether the system should continue, stop, retry or escalate

The ontology, capture protocol, annotation and acceptance plan follow that decision.

Start with the data bottleneck.

YPAI can run one defined stage or the complete data operation.

  1. Entry point: Collect new demonstrations

    Start here when: You need new human or robot interaction data.

    Typical output Structured demonstrations with synchronized observations, actions, states, outcomes and rights records.

  2. Entry point: Structure existing robot logs

    Start here when: You have data but cannot use it consistently across training, evaluation or review.

    Typical output Versioned episodes with normalized schemas, validated stream alignment and reproducible exports.

  3. Entry point: Annotate and evaluate

    Start here when: You need structured task meaning, outcomes or industrial-vision ground truth.

    Typical output Annotated episodes or visual datasets with ontology, reviewer records, QA results and acceptance status.

  4. Entry point: Build the data engine

    Start here when: You need operational infrastructure around the dataset.

    Typical output A controlled data and evaluation workflow connected to your model-development environment.

A robotic lawn mower on wet grass at night beside rocks, long grass and a boundary cable, one garden light

OUTDOOR AUTONOMY

The edge is part of the task.

A route learned on clean grass is not enough. Wet ground, stones, cable and the change from lawn to border all alter the next move.

Approach  →  boundary changes  →  route recovered

The useful record keeps the context that changes the route, not only the seconds when everything works.

Every episode should explain what happened.

The exact schema is defined for the customer's embodiment, policy stack and training objective.

A robot-learning episode can include the following layers.

Representative episode schema

Task identity
  • episode ID
  • task family
  • task variant
  • environment
  • scene
  • robot or embodiment
  • operator or collection method
  • protocol version
  • start and end conditions
Instruction and objective
  • natural-language instruction
  • structured task goal
  • target object
  • destination
  • constraints
  • prohibited actions
  • success definition
  • failure definition
Observations
  • fixed-camera video
  • egocentric video
  • wrist-camera video
  • multi-camera streams
  • RGB
  • depth
  • point cloud
  • audio
  • tactile or force data where available
  • environmental and device metadata
Robot and environment state
  • joint state
  • end-effector pose
  • gripper state
  • base pose
  • object state
  • tool state
  • scene state
  • sensor state
  • safety state
  • controller state
Actions
  • translation
  • rotation
  • gripper commands
  • joint commands
  • navigation actions
  • tool use
  • control mode
  • intervention
  • stop or retry action
Task structure
  • task phase
  • subtask
  • action segment
  • prerequisite
  • transition
  • contact event
  • state change
  • completion event
Outcome
  • successful
  • partially successful
  • failed
  • interrupted
  • unsafe
  • invalid
  • human intervention required
  • retry initiated
  • recovered
  • abandoned
Failure and recovery
  • failure point
  • failure category
  • preceding state
  • attempted action
  • environmental cause
  • perception error
  • planning error
  • control error
  • operator intervention
  • recovery action
  • final disposition
Provenance and quality
  • source
  • rights status
  • collection protocol
  • calibration record
  • sensor synchronization status
  • annotation version
  • reviewer status
  • issue history
  • acceptance status
  • dataset version

The episode schema is frozen before production scale.

The delivery is built around the robot task, capture setup and evaluation, not a generic image-labelling export.

A compact robot arm on a low wooden table in a domestic living room, positioned above a small block on a tray.
Capture

Capture the task
where it actually
happens.

The capture architecture depends on the model, embodiment, task and deployment environment.

A wheeled delivery robot crossing wet pavement beside a kerb in a residential street.
Robot-native execution
A warehouse aisle with racking and a worker in a high-visibility jacket walking away from the camera.
Industrial field collection

Label the interaction,
not only the object.

Physical AI annotation must preserve the relationship between time, state, action and outcome.

Close view of a robot gripper closing on a small wooden block resting in a shallow dish.
  1. Approach
  2. Contact
  3. Grasp
  4. Transfer
  5. Release
Object state / empty to filledOutcome / successFailure / dropped objectRecovery / successful retry

The taxonomy is built against the customer's model objective and operating environment.

Industrial vision data should represent the decision on the line.

A defect label without process context may not be enough to train or evaluate the system that must act on it.

Overhead view of a warehouse floor with stacked pallets, a worker in a high-visibility vest, and an autonomous mobile robot carrying a loaded pallet.
The same warehouse frame under strong glare.
Glare
The same warehouse frame in low light.
Low light
The same warehouse frame with motion blur.
Motion blur
The same warehouse frame partly occluded.
Occlusion

The final dataset records not only the label, but also the conditions under which the inspection decision was made.

Real data grounds the system.
Synthetic data expands the coverage.

Real episodes, simulation output and synthetic scenarios pass through one documented quality and acceptance process.

A grid of twelve dark panels, each holding the same sculptural form photographed at a different orientation. Real baseline Synthetic coverage
  1. Baseline
  2. Coverage gaps
  3. Generate
  4. Validate plausibility
  5. Evaluate against real
  6. Close the loop

Synthetic data is a coverage tool, not a substitute for real-world acceptance.

A single sculptural specimen mounted on a dark panel.
Episode-level acceptance

Quality follows the complete episode.

The quality process begins before collection or annotation and ends only when the dataset version is accepted.

AcceptedReworkRe-recordedReplacedRejected

Where accuracy or pass rate is the appropriate metric, YPAI typically proposes a 95% to 99% target range after calibration, with the exact metric, threshold and review method defined in the SOW.

Ships with the accepted version  /  episode manifest · sensor & stream schema · calibration package · ontology & annotation guide · quality report · provenance & rights record · dataset version record · checksums

Delivered in or mapped to RLDS, LeRobotDataset, MCAP, Parquet, JSONL, COCO, YOLO, KITTI or a customer-native schema. Format support is confirmed against the actual data, loader, toolchain and delivery requirement.

Different systems need different data operations.  Robot foundation model · manipulation and humanoid · industrial inspection · sensor and device · existing-data recovery.

Send the robot, task and data requirement.

The system, the task, the data source, the sensors, the scale, the annotation and the delivery format. YPAI returns feasibility, method, schema, QA and acceptance, delivery format and the open decisions before launch.

Related

Scope a Physical AI data project

The form adapts to the work, asks only for relevant details and sends your brief to the person who can act on it.

Service required

Your selection routes the brief to the right person.

A named project lead reviews every enquiry and replies within one business day