Fleet Ops & Safety
Incident Logging
L4 Cockpit HUD
ODD Coverage
How might we?
trust an autonomous test fleet enough to scale it, without losing the edge cases only a safety driver ever sees?
Platforms
Flutter cockpit HUD (iOS/Android) + React fleet dashboard (web)
Category
Autonomous Fleet Operations
Role
End-to-end: cockpit UI, data taxonomy, office dashboard
Design System
One taxonomy, six primitives
Every surface — cockpit, hazard map, office dashboard — reads from the same six primitives, styled consistently wherever they appear.
Severity
SEV1 · Critical Takeover
Subsystem
Motion Planner
Autonomy State
L4 AUTONOMOUS
ODD Coverage
Ingestion
Driver Monitoring
Gaze On Road
Severity & Status Colors
Critical (SEV1)
Hesitation (SEV2)
Comfort (SEV3)
Synced / OK
Neutral / Surface
Why It Works
Fewer blind spots.
One taxonomy from cockpit to office means no translation loss between what a driver saw and what an engineer reviews — the pattern autofleet’s safety-driver app and disengagement metrics point at.
Faster triage.
Severity and subsystem are tagged at the moment of the event, not reconstructed later — the edge-case and video-snippet tagging digitaldividedata builds for AV footage retrieval.
Provable ODD coverage.
Day/night/rain/traffic pass tracking against campaign targets makes exposure auditable, not anecdotal — operational exposure analysis and compliance/audit automation, both flagged across the research.
Approach & Process
End-to-End Ownership
Owned end-to-end design for Fleet Co-Pilot, spanning the cockpit HUD, the incident taxonomy, and the office review dashboard that shares it.
Reusable Incident Taxonomy
One classification drives three surfaces.
Subsystem × severity (perception, motion planner, localization, hardware/sensor, ODD boundary × SEV1–3) powers the driver’s 1-tap tags, the hazard map, and the office incident log — no re-mapping between tools.
Cross-Functional Partnership
Grounded every decision in how autonomy and fleet-ops teams actually work:
Safety drivers need sub-second quick-tags, not forms
Ops and AV engineers need exposure and disengagement analytics after the fact
Compliance needs an audit trail, not tribal knowledge
Multi-Fidelity Prototyping
Proved the taxonomy before building a backend for it:
Cockpit fixture data first, labeled simulated until a real ingestion pipeline exists
The office dashboard built as a second, independent consumer of the same schema
Validated the taxonomy held up outside the vehicle, not just inside it
Independent Research & Validation
Surveyed how autonomous-fleet operators actually run this today:
Route/dispatch optimization and predictive maintenance
Multi-vendor AV monitoring and safety-driver incident apps
Operational exposure analysis and edge-case/video tagging
1-Tap Incident Tagging
Sub-second logging, zero forms
Phantom Brake, Planner Freeze, Sensor Drop: three quick-tag buttons that log a subsystem and severity the instant something feels off, without pulling a driver’s eyes off the road.
The same pattern autofleet’s research flags as essential for safety-driver apps: incident reporting has to be faster than the incident.
Phantom Brake
SEV3 · Comfort
Planner Freeze
SEV2 · Hesitation
Sensor Drop
SEV2 · Hesitation
Live Telemetry
Everything a safety driver sees, structured
Autonomy state, jerk, lateral g, time-to-collision, and headway gap sit next to the driver-monitoring read: gaze, drowsiness score, hands-on-wheel, and takeover reaction latency.
Nothing here is inferred after the fact — it’s the same read the safety driver had, timestamped.
L4 AUTONOMOUS
Autonomy State
6.4s
Time to Collision
32 m
Headway Gap
0.12 g
Lateral Acceleration
Gaze On Road
Driver Monitoring
0.08
Drowsiness Score
Yes
Hands on Wheel
850 ms
Takeover Latency
Hazard Route Intelligence
Every disengagement becomes a place on the map
Fault and disengagement hotspots are geotagged with road segment, heading, and prior-disengagement count, then surfaced as an upcoming-hazard countdown as the vehicle approaches them again.
A tunnel that lost GPS once is a place the whole fleet gets warned about next time.
ODD Coverage & Ingestion
Provable exposure, not anecdotal
Day, night, rain/wet, and dense-traffic passes track against a per-route campaign target, so ODD coverage is a number, not a feeling.
Write rate, storage used, and sensor-stream sync status ride alongside it, so nothing gets logged in the cockpit and lost before it reaches the office.
Downtown ODD Loop — 6 / 10 passes
Day — 4 passes
Night — 1 pass
Rain/Wet — 1 pass
Dense Traffic — 2 passes
Payload Ingestion
148.4 MB/s write rate
842.5 / 1024 GB storage used
Sensor streams synced
