A fast score can hide a bad state
Drifting asks a small robot to estimate traction, motion, clearance, and route intent at once. A quick lap can still include stale evidence, wall contact, or an unrecoverable pose.
POC 02Embodied robotics
Kairos learns the dynamics.
Permission stays separate.
Kairos tests whether a small vehicle can learn drifting behavior from incomplete, mounted-sensor-style evidence while keeping maneuver intent, fast control, recovery, and physical authority in separate layers.
Simulation semantic envelopeSustain the arc
- Evidence
- Mounted fusion
- Gate
- Permission first
- Recovery
- May preempt
- Robot command
- None
Authority ceilingC0 / M0 · simulation evidence only
Simulator profile920 g RMX 2.5
Control architectureStrategy above fast control
Physical authorityNo implicit command path
Kairos · decision proof
Pressure Evidence Gate Result Trace
Read the simulator evidence in one bounded sequence: the motion pressure, the evidence available then, the permission gate, the semantic result, and the retained trace.
- Pressure
Nonlinear simulated motion under changing grip.
- Evidence
Distinct mounted-sensor-style streams and simulator telemetry.
- Gate
Permission and recovery constraints apply before preference.
- Result
A bounded phase and target envelope; fast control stays separate.
- Trace
Scenario, telemetry, outcome, and retained or rejected learning.
BoundarySimulation evidence only; not physical readiness, actuator authority, autonomous driving, or field deployment.
The pressure testFast, nonlinear motion.
Fast, nonlinear motion.
Incomplete evidence.
Intent above control, gates before ranking
Kairos considers a finite maneuver field, removes ineligible paths, and emits a bounded phase and target envelope. A separate fast controller owns steering and throttle.
Learning can be rejected and rolled back
Missing evidence remains missing. Faulted outcomes cannot seed a promoted profile. Safety and protected recovery can preempt strategy, and rejected learning intervals are restored.
Observed simulator sequenceReset the learner.
Reset the learner.
Change the grip live.
This continuous 1:53 simulator capture begins with retained learning already present, then shows an in-app learning-state reset from 56/56 to 0/0. The first post-reset run begins around 0:20 and reaches cycle 1/1 around 0:30. Later, the surface and tire compound change while the simulation remains active.
The reset is visible—and bounded to the app
The UI changes from 56/56 retained cycles to 0/0, then shows active drift telemetry and reaches 1/1. This supports an in-app reset sequence; storage erasure, a new process, and the baseline controller's contribution remain unresolved.
Surface and tires change without stopping the run
Both dropdown changes occur while motion and telemetry update. The capture then continues through 17/17 with no visible pause, crash, rollback, fault, or second reset.
The drift shallows without a visible loss of control
Displayed controlled-drift evidence falls from about 97% before the surface change to roughly 81–82% afterward and 77% after the tire change. In this simulator that reflects fewer signs of deep-angle drift, not a stability score; the vehicle continues without a visible spin, contact, pause, fault, or rollback.
Observed
One continuous recording shows an in-app reset from 56/56 to 0/0, post-reset cycle 1/1, two live simulator configuration changes, and continued motion through 17/17.
Inferred
The learner/controller stack appears to preserve stable, workable drift while adjusting drift depth across both changes in this run. The video does not isolate new learning from baseline heuristic control.
Unresolved
A formal stability metric, repeatability, seed and ordering effects, failure rate, exact reset semantics, causal adaptation, physical equivalence, and actuator authority remain unresolved.
Digital drift sandboxSee the world the
See the world the
controller is allowed to see.
Camera geometry, LiDAR, corner ranging, optical flow, and IMU remain distinct evidence streams. Simulator ground truth is not allowed to masquerade as a live physical observation.
Read the interface as four instruments
The continuous recording establishes the sequence. These focused panels make the setup, digital domain, and telemetry readable; each is one shown simulator state and does not independently establish change timing, causation, or formal stability.
01 · Experiment envelope
Hold the setup still before judging the learner
Shown simulator state: Foundation skidpad, Rubber mat, Hard plastic low grip, RMX balanced alignment, and Browser heuristic learner. The recording—not this still—shows when the surface and tire compound change.
02 · Digital domain
See the maneuver and the evidence envelope together
This paused, shown simulator state preserves the course and path history. Its status rail is intentionally excluded: in-app license and gate labels govern simulator maneuvers only—not deployment authority or physical actuation.
03 · Control and sensing
Requested motion, applied control, and body response diverge
This paused, shown simulator state displays cycle 19/19 beside requested control, tire use, fused perception, and IMU response. The values are session telemetry, not an external benchmark or evidence of physical actuation.
04 · Retained learning evidence
Ask which signal improved—and which one stayed flat
Judge, controlled drift, and target hold remain separate across the retained window. That makes plateaus, contradictions, and reward shortcuts visible instead of compressing every outcome into one score.
Architecture seeds
Six ideas future developers can carry into the next build.
Freshness and confidence remain attached to each sensor role.
Build, initiate, catch, sustain, transition, settle, or recover.
A score cannot compensate for a failed hard condition.
Strategy proposes a phase goal—not direct actuator output.
Fast control and recovery keep downstream authority.
Only complete, valid outcomes may shape retained memory.
Inspectable adaptationA score is one signal.
A score is one signal.
Not the whole result.
The sandbox retains judge, controlled-drift, target-hold, fault, context, prediction, and rollback evidence separately. More cycles do not automatically become improvement.
Historical stored simulator replay12 / 12 course gates · 8.83 s lap · 6.3 s controlled drift
Judge average 55.5 · phase coverage 75% · settle phase absent. One tracked replay with a visible gap—not a current benchmark.
Different result shapeNot one policy
Not one policy
chasing one reward.
Kairos’s proof is the separation: perception, intent, control, safety, outcome, and learning remain reconstructable layers.
01SenseTyped mounted evidence
→02GatePermission before rank
→03ProposeSemantic target envelope
→04ControlFast loop retains simulated control
→05ReviewRetain, reject, or recover
Evidence boundary
Simulation is not the physical robot.
Kairos remains at the C0 / M0 authority ceiling. The material shown here is simulation evidence only; it does not establish semantic or implementation conformance, trace demonstration, physical equivalence, autonomous drifting, deployment readiness, or permission to actuate the robot.
The simulator and live-monitor views expose no implicit hardware command path. Physical familiarization remains a separately gated, supervised commissioning domain.