Systems and records
World model physical units validation: catching impossible examples
By SourceX Editorial · Updated
10 sections
Short answer
Detect physically impossible examples caused by wrong units by declaring every field's unit, frame and time base from source documentation, then recomputing quantities a second way: speed from positions, gravity at rest, rates from timestamps, and limits from equipment specs. Errors that sit at a fixed conversion ratio across an export are units problems; convert them once with provenance, label real extremes and quarantine what stays unexplained.
Video and sensors
World model physical units
Key takeaways
- A wrong unit produces a consistent, plausible-looking dataset with the wrong physics, not an obvious outlier.
- Recompute each quantity a second way; a mismatch at a fixed conversion ratio is the signature of a units error.
- A stationary accelerometer near 9.81 m/s^2 is correct; a reading near 1 means the data is in g.
- Separate units errors from sensor faults, gaps, real extremes and unlogged interventions before changing anything.
- Convert with provenance, label real extremes, quarantine the unexplained and never clip to plausible values.
How do you detect physically impossible examples caused by incorrect units?
Declare the unit, coordinate frame and time base of every numeric field from the source system's own documentation, then recompute each quantity a second way and compare. Speed derived from position and time should match reported speed, a stationary accelerometer should read about one g, and nothing should exceed what the equipment can physically do. A mismatch that sits at a known conversion ratio, such as 2.237 between miles per hour and meters per second, points to a units error rather than a real event.
World model physical units validation is that process run before training, on every source and every export version. The checklist below makes it repeatable, and the triage section explains how to tell a units error from a sensor fault, a missing observation or a genuinely extreme event.
Why does a wrong unit produce an example that looks fine but is impossible?
A units error rarely produces an obvious outlier; it produces a consistent, plausible-looking world with the wrong physics. A forklift whose speed in miles per hour is stored as meters per second travels more than twice as fast as it did, accelerates harder and stops in less distance, and every record from that export agrees with the error.
The classic warning is NASA's Mars Climate Orbiter. Its mishap board found the root cause was the failure to use metric units in a ground software file, where thruster performance data in English units was used instead of metric; the small errors accumulated until the trajectory was about 170 kilometers lower than planned.
World models are sensitive to this because they learn consequences from examples. Kang and colleagues report that video generation models trained on simulated motion generalize well in distribution but fail out of distribution and behave in a case-based way, mimicking the closest training example rather than abstracting the law. Training cases with the wrong physics are therefore not averaged away; they become cases the model can imitate.
The units and physical constraint validation checklist
Run the checks in order, because later checks assume earlier ones passed. Record the result per source and per export version, not once for the whole dataset.
| Check | What it catches | How to test | Evidence to keep |
|---|---|---|---|
| 1. Unit declared per field | Fields with no stated unit or a guessed one | Map each field to a machine-readable unit code from the source documentation | Unit dictionary with the documentation page and date |
| 2. Symbols, not abbreviations | Ambiguous labels such as sec, hrs or g | Reject free-text units; resolve g as gram or standard gravity explicitly | List of resolved ambiguities |
| 3. Time base | Milliseconds read as seconds, mixed clocks, daylight saving jumps | Timestamps monotonic per device; median step matches the stated sample rate | Clock source and time zone per device |
| 4. Frame and handedness | Left-handed frames, swapped axes, wrong origin | Known motion (driving forward, lifting) moves along the expected axis and sign | Frame description per sensor |
| 5. Stationary sanity | Accelerometers in g, gyros in degrees | At rest, acceleration magnitude near 9.81 m/s^2 and angular rate near zero | Rest segments used for the test |
| 6. Cross-derivation | Speed or rate fields that disagree with positions | Recompute velocity from position deltas over time and compare the ratio | Ratio histogram per source |
| 7. Capacity limits | Values beyond the equipment's rated speed, load or reach | Compare against manufacturer specifications for that model | Spec sheet reference and model list |
| 8. State constraints | Negative stock, mass appearing from nowhere, overlapping occupancy | Balance inputs and outputs per step and per location | Constraint list and violation counts |
| 9. Conversion provenance | Hand-typed or rounded conversion factors | Every conversion uses a factor from a published table, applied once | Factor source and pipeline version |
| 10. Intervention log | Manual overrides and resets that break continuity | Join operator actions and maintenance events to the time series | Intervention table linked by time and device |
Which conventions should the unit dictionary follow?
Pick one documented convention and convert into it once, at ingestion. Robotics software has a widely used default: ROS REP 103 standardizes on SI units (meters, kilograms, seconds, radians) and right-handed frames with x forward, y left and z up for a body, and east-north-up for short-range geographic frames.
Sensor messages make the same choice explicit. The ROS sensor_msgs Imu message says accelerations should be in m/s^2, not in g's, and rotational velocity in rad/sec, and REP 145 says an accelerometer at rest reports the specific force due to gravity, so an upward z axis should read +g. That is why a stationary reading near 9.81 is a pass, not an error.
For written unit labels, the NIST Guide to the SI treats unit symbols as mathematical entities rather than abbreviations and gives sec, mins and hrs as forms that are not permitted. Where units travel between systems as text, a code system such as UCUM, designed for unambiguous machine-to-machine communication of units, removes the guesswork.
Which unit mistakes turn up most in operational records?
Most errors come from exports and joins, not from sensors. The table lists patterns worth testing in any fleet, warehouse, factory or field-service dataset, with the quick detector for each.
Joins deserve their own attention because world-model episodes usually combine a fast sensor stream with slow business events. A task record that says a pallet was put away at a given minute, joined to telemetry sampled many times a second, can attach the consequence to the wrong action if the two systems use different clocks or time zones. Test the join by checking that each business event lands while the vehicle or robot is physically at the right location, and measure the offset rather than assuming zero.
Units can also change over time inside one source. A telematics firmware update, a new export template or a vendor migration can switch a field's unit or precision partway through the history, so run the checks per export version and per month, not only per source.
| Mistake | How it shows up | Quick detector |
|---|---|---|
| Miles per hour stored as meters per second | Vehicles move about 2.24 times faster than positions allow | Derived to reported speed ratio near 0.447 or 2.237 |
| Milliseconds read as seconds | Events stretched a thousandfold or rates near zero | Sample interval of 1,000 where 1 is expected |
| Degrees read as radians | Spins of dozens of turns between frames | Rotation between frames about 57.3 times what position or heading changes allow |
| Acceleration in g treated as m/s^2 | Rest magnitude near 1 instead of 9.81 | Stationary segment check |
| Pounds and kilograms mixed in payload | Loads above rated capacity at one site only | Per-site payload distribution against the spec sheet |
| Millimeters from CAD mixed with meters from tracking | Objects a thousand times too large or too small | Bounding size against known part dimensions |
| Local time without zone | Duplicate or missing hour twice a year | Gaps or overlaps at daylight saving transitions |
| Cumulative counters treated as rates | Ever-growing values that look like acceleration | Monotonic increase with resets at power cycles |
How do you tell a units error from a sensor fault, a gap or a real extreme?
Before converting anything, decide which of five explanations fits, because each needs a different fix. A units error is systematic and sits at a known ratio across a whole export; a sensor fault is local in time; a missing observation shows up as a gap; a real extreme event has physical corroboration; and an unlogged intervention coincides with a human action.
| Explanation | Signature in the data | Test that separates it | Usual handling |
|---|---|---|---|
| Units or conversion error | Constant ratio across a source or export version | Ratio clusters at a published conversion factor | Convert once with provenance and re-run all checks |
| Sensor fault or calibration drift | Errors limited to one device or time window | Compare with a neighboring sensor or calibration log | Flag the window; keep the device history |
| Missing observations | Large time step makes derived speed spike or vanish | Sample interval against stated rate | Mark the gap; do not interpolate dynamics across it |
| Real extreme event | Short spike with corroborating effects | Damage report, maintenance ticket or second sensor | Keep and label; these are valuable consequence examples |
| Unlogged intervention | Discontinuity at an operator action | Join to operator and maintenance logs | Split the episode at the intervention |
Worked example: illustrative forklift telemetry joined to task records
Illustrative only. A buyer assembles world-model training episodes from forklift telemetry at three warehouses, joined to warehouse management task records that say which pallet moved where. Each episode pairs the vehicle state and operator action with the consequence: the pallet arrives, a rack position fills or a task is reassigned.
Check 6 fails at one site: speed derived from positions is about 0.45 of the reported speed on every vehicle, and the ratio does not drift. The vendor export documentation for that site's telematics version lists speed in miles per hour while the other two list meters per second. The team converts with the published factor, records the export version, and re-runs checks 6 and 7; the site now matches the others.
Check 7 then flags short vertical acceleration spikes above what the team expected. Maintenance tickets show they cluster at one dock plate with a known lip, so they are real impacts. They stay in the dataset with a label, because a world model that never sees a bump has not learned the warehouse. A second site's timestamps advance by about 1,000 per sample; its device clock reports milliseconds, and conversion fixes it.
The episodes that cannot be explained, roughly 2% in this illustration, are quarantined with the reason rather than clipped to plausible values.
Should a failing record be converted, flagged or quarantined?
Convert only when the error is proven systematic and the correct unit is documented by the source; flag when the value is real but unusual; quarantine when the record is impossible and no explanation survives triage. Never clip impossible values to a plausible range, because clipping invents physics the system never showed.
At a glance
Should a failing record be converted, flagged or quarantined
Every change should be reversible. Keep the original value, the applied factor, the check that triggered it and the export version, so a later reviewer can rebuild the episode and a buyer can see how many records each rule touched.
- Convert: constant ratio, documented unit, all downstream checks pass after conversion.
- Flag: corroborated extreme event, recorded intervention or known calibration window.
- Quarantine: impossible state with no corroboration, mixed units inside one record, or an undocumented frame.
- Escalate: any rule that touches a large share of one source, which suggests the source documentation is wrong.
Where do automated unit checks run out?
The checklist finds inconsistencies; it cannot decide what the physics of a specific operation should be. Capacity limits need the model-specific specification sheet, state constraints need someone who knows how the site actually works, and sensor fusion outputs may already mix frames in ways only the integrator can explain.
Bring in an operating domain expert and a world model researcher when constraint violations cluster by site, shift or equipment model, when the source cannot document its units or frames, or when a correction would change a large share of episodes. Those are signs the dataset's description of reality is wrong, not just individual records.
What should a request for operational records say about units?
Most of the checklist is cheap when the supplier sends a unit dictionary, frame descriptions, clock sources, export versions and calibration dates with the data, and expensive when the buyer reconstructs them afterward. Put those items in the request, along with the right to ask follow-up questions about the source systems.
Telemetry and task records are sourced for a specific buyer: SourceX has no inventory and approaches U.S. businesses once a request describes the scope, volume, language, format, timing and licensing needs, then handles qualification, rights review, contracting and preparation with the supplying company. Ownership and licensing rights to telemetry are a separate question from data quality and are settled in that review.
Frequently asked questions
What is world model physical units validation?
It is the set of checks that confirm every numeric field in world-model training data has a declared unit, frame and time base, and that the resulting states obey basic physical and operational constraints. It runs before training, per source and per export version.
Why does a stationary accelerometer read about 9.81 m/s^2?
An accelerometer measures specific force, so at rest it reports the force due to gravity. ROS REP 145 says a sensor at rest with its z axis up should indicate +g, which is why a reading near 1 in that situation suggests the data is in g, not m/s^2.
Should impossible values be clipped to a realistic range?
No. Clipping hides the cause and teaches the model dynamics that never happened. Convert when the unit error is proven, keep and label real extremes, and quarantine what cannot be explained.
How can I tell a unit error from a sensor fault?
A unit error is systematic: the same ratio across an entire source or export version, usually at a published conversion factor. A sensor fault is local to one device or time window and disappears when compared with a neighboring sensor.
Which unit convention should a robotics dataset use?
Any documented convention works if it is applied once and consistently, but many robotics teams use ROS REP 103: SI units, radians and right-handed frames with x forward, y left and z up. Record the choice in the dataset documentation.
Sources
- NASA's Mars Climate Orbiter mishap board found the root cause was English-unit thruster data used where metric units were required in ground software, leaving the trajectory about 170 kilometers lower than planned. Source
- Kang et al. report that video generation models trained on simulated physics generalize in distribution but fail out of distribution and mimic the closest training example rather than abstracting physical laws. Source
- ROS REP 103 standardizes on SI units including meter, kilogram, second and radian, right-handed frames, x forward, y left and z up, and ENU for geographic frames. Source
- The ROS sensor_msgs Imu message specifies accelerations in m/s^2 rather than g's and rotational velocity in rad/sec. Source
- ROS REP 145 says an accelerometer at rest reports the specific force due to gravity, so an upward z axis indicates +g, in m/s^2. Source
- The NIST Guide to the SI says unit symbols are not abbreviations and gives sec, mins and hrs as forms that are not permitted. Source
- NIST's conversion table gives 4.4704 E-01 as the factor from mile per hour to meter per second. Source
- UCUM is a code system for unambiguous electronic, machine-to-machine communication of quantities and their units. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.