Cadence live
TOP VIEW METRES
Waiting for model state
Targets reached— / —
Motion prediction error—
Compute / step— ms
WHEEL COMMAND
L
—
R
—
BODY ADAPTATION EXPERIMENT 01
One wheel changes.
Learning continues.
A simulated rover follows targets. Weaken its right wheel and watch how learning changes its motion predictions and control.
00:00.0
SIMULATED TIME STEP 0
Waiting for simulator
Preparing the learned body models. The first checkpoint may take a moment.
01
Normal motionShared target schedule
02
Weaken one wheelLearn from measured motion
03
Restore the bodyProbe first, then relearn
RIGHT WHEEL—
Return probe: learning is disabled for every model. These measurements test the state retained after the weak wheel. Resume learning to measure relearning separately.
The same controller. Two learning conditions.
rover path ⊕ targetLEARNING ENABLED
Cadence frozen
TOP VIEW METRES
Waiting for model state
Targets reached— / —
Motion prediction error—
Compute / step— ms
WHEEL COMMAND
L
—
R
—
Cadence learns motion; the shared controller chooses actions. Every model receives commands and odometry. The wheel strength and change signal are hidden from the models.
Inside the live model
PATCH READBACKLocal settled states and prediction errors from the running Cadence patches.
Patch readback appears after initialization.
Bounded software patches read local state through ports, record motion, and use feedback to update the same patch rule.
Motion prediction error
RECORDED SAMPLESCadence liveCadence frozen
Lower is better. Error scale and history come from the simulator.
Technical comparisons
Same sensors · same action selection · separate experienceTargets reached by phase
REACHED / ATTEMPTEDThe return probe disables learning. Relearning starts from that same life. A dash means this phase has no recorded attempts.
| Model | Normal wheels | Weakened wheel | Return probe learning off | Return learning learning enabled |
|---|---|---|---|---|
| Waiting for measured trajectories. | ||||
Totals include the active target attempt. Each target has the same time budget. Frozen Cadence keeps its bootstrap state throughout.
Run record
SEED —- —
No recorded events.
Keep the evidence
REPRODUCIBLE LIFESave the full simulation state to resume the same life in this page, or export the measurements and experiment settings.
Restoring a checkpoint tests continuation. Return to normal mechanics measures performance with the learned state kept in place.
Experiment scope & measurement details +
This is a differential-drive simulation of a learned body model inside a supplied target controller. A single displayed run does not establish an advantage over adaptive or neural baselines, physical robot performance, or protected memory.
Targets reached count completed targets; attempts include target timeouts. Compute time includes model prediction, action selection and learning. The run record and receipt identify the protocol and phase measurements.
Every model and the rover body run in this page, in a worker beside the display. The Cadence brains use a JavaScript edition of the population solver of cadence-net 0.70.0, checked against the Python library. While the tab is hidden the life waits.
- Protocol
- Waiting for simulator