RAST
The Missing Layer in ADAS

Every driver optimizes
for something different

RAST extracts the driver's motor cost function. Not what they do, but what they optimize for. This enables ADAS that predicts, personalizes, and earns trust.

~9 min
Between L2 driver interventions (AAA 2025)
Majority
Cut-in events where drivers override ADAS
0
Systems personalizing to driver motor identity
Motor Cost Function

Two drivers. Same car.
Different intent.

Extracted from everyday driving: How each driver trades off smoothness, safety, efficiency, and responsiveness. Stable traits transfer across vehicles and conditions.

Driver A
The Smooth Operator
Optimizes for smooth trajectories and wide safety margins. Brakes early and gently. Gradual lane changes. Low rapid-correction practice.
Smoothness
0.88
Safety margin
0.82
Efficiency
0.45
Responsiveness
0.35
Stability
0.85
Smoothness - trait Safety margin - trait Responsiveness - state
Driver B
The Efficient Racer
Optimizes for time efficiency and responsive handling. Late braking, tight margins, decisive corrections. Strong rapid-correction repertoire.
Smoothness
0.35
Safety margin
0.40
Efficiency
0.90
Responsiveness
0.85
Stability
0.50
Efficiency - trait Responsiveness - trait Safety margin - state
Extract → Predict → Intervene

Generic ADAS solves the wrong problem.

A car cuts into the passing lane at 110 km/h. Generic ADAS applies the same response to every driver. RAST predicts each driver's specific failure mode, and designs the right intervention.

Highway cut-in, 110 km/h, passing lane
Emergency
Generic ADAS Driver failure RAST intervention
Driver A: Smooth Operator
↑ ↑ ↑ PASSING MIDDLE RIGHT AEB GENERIC ADAS Brakes for forward collision Doesn't predict the left swerve RESULT AEB fires, but driver panics → swerves into barrier RAST ADAS (−1.8s) Gentle brake before panic + ESC armed for left swerve CUT YOU
Generic ADAS
Driver panics
RAST ADAS
1
Extract - cost function from everyday driving
▾
w_smooth = 0.88 w_safety = 0.82 w_responsive = 0.35
Optimizes for gradual corrections. Low rapid-correction weight = unpracticed sharp maneuvers.
2
Predict - 38° swerve left toward barrier
▾
OFC forward solve under panic: Decision time collapses from 2s to 200ms. Motor system must execute a sharp correction it has never practiced. High smoothness + low responsiveness → overshoots into barrier zone.
Habit tracking cannot do this. The driver has never panicked before. Only the cost function, run forward under new constraints, produces this prediction.
3
RAST: Early gentle braking (−1.8s) + ESC armed
▾
Prevents panic onset entirely. The dangerous swerve never happens.
Generic ADAS fires standard AEB brakes for forward collision. But the real threat is the left swerve it doesn't anticipate. AEB fires, driver panics anyway, swerves into barrier. Solved the wrong problem.
Driver B: Efficient Racer
↑ ↑ ↑ PASSING MIDDLE RIGHT AEB GENERIC ADAS AEB fires early (−2.0s) Driver perceives as premature RESULT Driver overrides AEB, holds speed → brakes too late → rear-end RAST ADAS (−0.6s) Waits, fires hard AEB at last safe moment + lane guard CUT YOU
Generic ADAS
Driver overrides
RAST ADAS
1
Extract - cost function from everyday driving
▾
w_efficiency = 0.90 w_responsive = 0.85 w_safety = 0.40
Optimizes for time efficiency. Brakes late habitually. Low safety-margin weight.
2
Predict - brakes 400ms too late → rear-end + overrides early AEB
▾
OFC forward solve under emergency: High efficiency weight means motor system is trained to delay braking for speed. Under sudden cut-in, same optimization delays braking past safe threshold. And: this driver will override early intervention - perceives it as premature.
Habit tracking says "brakes late." RAST says why the efficiency cost function computes exactly how late, and predicts the override of premature intervention.
3
RAST: Late sharp AEB (−0.6s). Matches aggressive style
▾
Waits until last safe moment. This driver accepts hard braking. Matches aggressive style. Lane departure prevention active.
Generic ADAS fires AEB early (−2.0s). This driver perceives it as premature, overrides by accelerating. Neither human nor system has clean control. Intervention made it worse.
Control Handoff

The most dangerous
3 seconds in driving

When ADAS returns control, RAST shapes the vehicle state to match this specific driver. First action is continuation, not correction.

Today - Generic handoff
One-size-fits-all
Following distance
1.2s (default)
Mismatch
Lane position
Center
Mismatch
Speed trend
Decelerating
Mismatch
Notification
Beep + visual
Generic
Driver spends ~3 seconds correcting
Maximum cognitive load at worst moment
RAST - Personalized handoff
Shaped to Driver A
Following distance
2.1s (preference)
Match
Lane position
Slight left
Match
Speed trend
Stable
Match
Notification
Early + haptic
Personal
First action is continuation
Zero correction needed
Kinematic Prediction

Know how the movement ends
before it finishes

From thousands of everyday steering inputs, RAST learns the driver's kinematic signature. In 100ms it predicts the full trajectory.

0 ms
Steering onset. Angular acceleration detected. Direction: left.
100 ms
Profile matched. Velocity ramp = Driver A. Single-peaked, high-amplitude. Predicted: 38° left.
200 ms
Trajectory projected. Completion → exits lane in 1.2s. Decision: INTERVENE
350 ms
Correction applied. Limited to 22°. Vehicle stable. Natural feel preserved.
From everyday driving, thousands of inputs, RAST builds a kinematic signature that predicts emergency behavior before it completes. Not reaction. Prediction.
Contextual Motor Graph

The driver's motor passport

A computational motor identity - what they can do, how they move, why they move that way, and what transfers to new conditions.

◈ CMG: Driver A
Motor Identity Layers
From physical constraints to transferable intent.
🦴
Structural
Arm reach, seating position, grip strength, reaction baseline
〰️
Kinematic
Single-peaked velocity, smooth corrections, low trial variability
💪
Kinetic
Low peak force, gradual torque ramps, symmetric effort
⚡
Cost Function
w_smooth=0.88 ± 0.04 · w_safety=0.82 ± 0.06 · w_effort=0.45 ± 0.12
🏷️
Transfer Tags
TRAIT: smoothness, safety, correction style · STATE: effort, responsiveness
Versioned · Uncertain (confidence bounds) · Context-indexed
⬡ Why CMG is different
What systems lack today
No current system combines all elements.
📊
Driver monitoring
Measures alertness. Nothing about how or why the driver drives.
📈
Behavior profiling
Records patterns. Patterns don't generalize to new conditions.
🧠
CMG adds: the "why"
Extracts the optimization strategy. Generalizes across vehicles, roads, conditions.
🔄
CMG adds: trait vs. state
Stable traits transfer to any car. States adapt to fatigue, weather, road.
☁️
CMG adds: portability
Compact vector. New car reads it instantly. Seat memory for driving style.
Privacy-preserving Mathematical abstraction, not a recording. No routes, no timestamps.

RAST - Motor Intelligence Infrastructure

We don't learn what the driver does. We understand what the driver optimizes for.