Module 5 — Safety Standards and Deployment¶
Deep Dive · Specialization
Specialized course identity for Module 5 — Safety Standards and Deployment.
Parent: Phase 5 — Autonomous Driving
Time: 3–6 months
Prerequisites: Modules 1–2 (fundamentals + openpilot system understanding). Module 4 (Advanced Perception) is recommended but not required.
Why safety and deployment¶
A single perception accuracy percentage does not establish vehicle safety: the operational domain, failure modes, exposure, and system response determine the risk. This module covers the standards, testing methodologies, and deployment patterns that bridge the gap between a working prototype and a production ADAS.
1. Functional Safety Standards¶
-
ISO 26262 (Road Vehicles — Functional Safety):
- ASIL classification: A (lowest) through D (highest) safety integrity levels.
- Hazard Analysis and Risk Assessment (HARA): severity, exposure, controllability.
- Safety goals → functional safety requirements → technical safety requirements.
- Safety lifecycle: concept → development → production → operation.
- Hardware metrics: SPFM (Single Point Fault Metric), LFM (Latent Fault Metric).
-
SOTIF (ISO 21448 — Safety of the Intended Functionality):
- Addresses risks from sensor limitations, algorithm uncertainty, and unpredictable environments — distinct from malfunctioning E/E behavior addressed by ISO 26262, including hardware and software aspects.
- Known/unknown unsafe scenarios, triggering conditions.
- Validation strategy: reduce residual risk from intended functionality.
-
Safety architecture patterns:
- Redundancy: Dual-channel monitoring (e.g., two independent perception paths).
- Diverse redundancy: Different algorithms or sensors for the same safety function.
- Plausibility monitoring: Cross-check perception outputs against expected physical constraints.
- Graceful degradation: Fallback to simpler, safer behavior when primary system degrades.
Projects: * Perform a HARA for a lane-keeping assist function. Assign ASIL levels to identified safety goals and propose mitigations. * Design a safety architecture for an ACC system: identify single-point failures and propose redundancy/monitoring.
2. V2X (Vehicle-to-Everything) Communication¶
-
Communication standards:
- DSRC (Dedicated Short-Range Communications): 802.11p-based, mature but limited bandwidth.
- C-V2X (Cellular V2X): LTE-V2X (PC5 sidelink) and 5G NR-V2X; latency and reliability depend on radio mode, congestion, coverage, and deployment.
- V2V (vehicle-to-vehicle), V2I (vehicle-to-infrastructure), V2P (vehicle-to-pedestrian).
-
Cooperative perception:
- Vehicles share sensor data or object detections via V2X.
- Extends effective sensing range beyond individual vehicle FoV.
- Challenges: latency, bandwidth, data format standardization.
-
V2X security:
- IEEE 1609.2, ETSI ITS Security.
- Certificate authorities, pseudonymous authentication.
- Privacy-preserving communication (location privacy vs. safety).
3. ADAS Validation and Testing¶
-
Scenario-based testing:
- Scenario descriptions: ASAM OpenSCENARIO; simulation interfaces: ASAM OSI. These are standards/interfaces, not scenario databases.
- Systematic coverage: edge cases, SOTIF-relevant scenarios, ODD (Operational Design Domain) boundaries.
- Concrete vs. logical vs. functional scenarios.
-
Hardware-in-the-Loop (HIL) testing:
- Inject synthetic sensor data into production ECUs.
- Validate ADAS software under controlled, repeatable conditions.
- Closed-loop HIL: ECU outputs feed back into simulation.
-
Shadow mode deployment:
- Run experimental perception in parallel with production system — no actuation.
- Log disagreements between experimental and production outputs.
- Offline evaluation: curate targeted test sets from disagreements.
- Metric: disagreement rate, false positive/negative analysis.
-
Field operational tests (FOT):
- Controlled real-world testing with safety drivers.
- Data collection: metrics, edge cases, system performance under ODD.
- Regulatory requirements by region (US, EU, China).
Projects: * Start with a software-in-the-loop replay that injects synthetic frames into a perception process. Call the extension HIL only when a named physical target ECU participates, with documented I/O timing and simulator interfaces. * Compare an experimental perception algorithm with a baseline in offline replay or a local simulator. Collect disagreements and ground-truth outcomes; disagreement alone does not identify which model is correct. * Create an OpenSCENARIO scenario for a pedestrian crossing at an intersection. Run it in CARLA with your Module 1 controller.
Resources¶
| Resource | Why |
|---|---|
| ISO 26262 Standard | Foundational safety standard for automotive electronics |
| ISO 21448 (SOTIF) | Safety of intended functionality for ADAS/AD |
| Autonomous Vehicles and Functional Safety (Tier 1 guides) | Practical guides to applying ISO 26262 + SOTIF |
| 5GAA | C-V2X standards, use cases, deployment guidance |
| OpenSCENARIO | Scenario description standard for ADAS testing |
| CARLA | Simulation for HIL and scenario-based testing |
Next¶
→ Module 6 — Lauterbach TRACE32 Debug (optional) — In-circuit debug and trace for automotive ECUs.
Turn a failure scenario into a testable requirement¶
For a teaching lane-following system, distinguish: (a) a corrupted sensor message, (b) a valid image whose glare hides the lane, and (c) a stale but plausible action. Map each to the relevant malfunction, intended-function limitation, or timing/system requirement. Define the response and maximum detection time before testing it.
A HARA exercise produces a reasoned assignment for a safety goal, not an ASIL label for an entire neural network. Keep the evidence chain explicit: hazard → goal → requirement → monitor/fallback → test result. ISO 26262 and ISO 21448 have different scopes; consult the actual project-applicable standards for compliance work. ISO 26262 scope, ISO 21448 scope.
Use replay or simulation for the baseline, and call a test HIL only when real target hardware participates in the loop. A classroom simulation report does not certify a road vehicle. Pass the monitor's timestamps and failure trace to Module 6 or the SoC controller testbench.
Worked available artifact: use the executed delayed-observation scenario. The teaching requirement rejects acquisition age above 150 ms; a 200 ms delay produces 156 stale observations after four missing ticks. Its chosen fallback brings speed to zero, but the lane-following task fails because lateral error remains 1 m. Record detection/response behavior and task outcome separately. This exercises one stated requirement; it does not establish that the fallback is appropriate for a real road situation.
Practice and assessment¶
Entry diagnostic (prerequisite check): Distinguish a hazard, a safety requirement, a test result, and certification evidence. If this is unfamiliar, use the prerequisite material above or the roadmap diagnostics before starting the lab.
- Baseline lab: Create a traceable hazard-to-test example for a simulated vehicle function.
- Guided variation: Replay an explicit sensor or timing fault and test the specified fallback.
- Independent transfer: repeat the method on a new input or configuration and explain where your original assumptions stop applying.
Evidence to submit: Requirement/test mapping, scenario outcomes, timing, assumptions, and unresolved evidence; no certification claim. Keep raw results and reproduction steps; distinguish measured results from estimates or work not run.
Self-check: Which observation would disprove your explanation of the result? Identify a competing explanation and the smallest experiment that separates them. A passing answer names a concrete predicted difference, not only another tool to try.
Exit criteria: reproduce the baseline, demonstrate the negative/variation check, and meet the shared completion rubric. Use the lesson cycle for retrieval practice and spaced review.