VLM/VLA Reliability Benchmarks for Autonomous Driving

Key Takeaway: Driving VLMs and VLAs are useful only when their visual grounding, corruption robustness, text sensitivity, and action validity are measured explicitly. Current benchmarks such as DriveBench, CODA-LM, DriveLM, and RoboDriveBench show that models can produce plausible but visually ungrounded answers, degrade under sensor corruption, and exploit text priors. For airside autonomy, a VLM/VLA benchmark must include text-only controls, wrong-context prompts, sensor corruptions, operational-rule questions, and closed-loop action checks.

---

Reliability Question

VLM/VLA driving systems promise semantic reasoning, instruction following, explanations, and direct trajectory generation. The safety question is not "can the model answer driving questions?" but:

• Did it look at the right visual evidence?
• Did it ignore misleading text priors?
• Did it remain consistent under corruption, missing frames, and prompt errors?
• Did its explanation match the action it proposed?
• Did the proposed action obey the domain rules and remain controllable?

These questions require reliability benchmarks beyond standard VQA or open-loop planning scores.

---

Benchmark Taxonomy

Benchmark class	What it measures	Example benchmarks
Driving VQA and reasoning	Scene, prediction, planning, behavior, explanation QA	DriveLM, LingoQA, nuScenes-QA
Corner-case VLM evaluation	Rare or safety-critical driving scenes	CODA-LM
Visual grounding and hallucination	Whether answers depend on actual image/video evidence	DriveBench text-only and corruption settings
Sensor/prompt robustness	Performance under weather, sensor, motion, transmission, and prompt corruption	DriveBench, RoboDriveBench
VLA trajectory/action validity	Whether VLM/VLA outputs safe trajectories or controls	RoboDriveBench, LMDrive/Bench2Drive-style closed-loop
Explanation-action consistency	Whether rationale matches planned action and scene facts	DriveLM, DriveBench, CoT-style rubrics
Domain-specific operational reasoning	Rules and procedures for the operating domain	Airside benchmark extension needed

---

Key Benchmarks and Lessons

DriveBench

DriveBench evaluates VLM reliability across clean, corrupted, and text-only settings. It covers perception, prediction, planning, and behavior/explanation tasks with 19,200 frames and 20,498 QA pairs. Its most important lesson is that VLMs can generate plausible driving responses from text cues or dataset priors even when visual evidence is degraded or missing.

Recommended use:

• Use text-only controls to estimate language-prior leakage.
• Report clean/corrupted/text-only deltas, not only absolute scores.
• Include multiple question types: multiple choice, open-ended, and visual grounding.
• Track whether a model localizes the relevant object or simply answers generically.

CODA-LM

CODA-LM targets self-driving corner cases and provides automated evaluation for large VLMs on general perception, regional perception, and driving suggestions. It is valuable because corner-case evaluation is closer to safety-critical deployment than average-case captioning.

Recommended use:

• Build domain-specific corner-case suites rather than only normal scenes.
• Separate global scene understanding from regional/localized perception.
• Use evaluator models carefully and calibrate them against human judgment.

DriveLM

DriveLM frames driving as graph visual question answering. It connects perception, prediction, planning, behavior, and motion tasks through human-written reasoning logic. It is useful for evaluating whether a model can preserve structured dependencies from "what is present" to "what will happen" to "what should ego do."

Recommended use:

• Use graph dependencies to detect inconsistent answers.
• Keep questions tied to object IDs, spatial relationships, and future intent.
• Evaluate planning explanations separately from final action quality.

RoboDriveBench / RoboDriveVLM

RoboDriveBench focuses on robust VLM-based end-to-end trajectory prediction. It introduces simulated sensor corruption and prompt corruption categories and evaluates trajectory prediction cases at scale. Its core lesson is that VLA reliability must be measured in action space, not only language space.

Recommended use:

• Include both sensor and prompt corruption.
• Measure trajectory error, collision risk, and rule compliance.
• Test cross-modal recovery using LiDAR/radar where available.

---

Reliability Metrics

Metric	Definition	Why it matters
Clean score	Task score on uncorrupted input	Baseline capability
Corruption delta	Clean minus corrupted score	Robustness to weather, blur, missing frames, compression, glare
Text-only score	Score without visual input	Detects language-prior leakage
Grounding accuracy	Whether cited object/region matches the answer	Prevents plausible but ungrounded explanations
Hallucination rate	References to nonexistent objects, signals, people, or hazards	Safety-critical for VLM co-pilots
Consistency score	Agreement among perception, prediction, planning, and explanation answers	Detects broken reasoning chains
Action validity	Trajectory/control obeys constraints	Required for VLA deployment
Explanation-action alignment	Explanation supports the selected action	Useful for incident review and trust
Calibration / abstention	Confidence matches correctness and model can say unknown	Needed for fallback gating
Deadline miss rate	Output arrives after planner deadline	Reasoning that is too late is not usable

---

Relevance by Domain

Generic Road AV

Driving VLMs can help classify rare semantic hazards, explain decisions, and support data mining. They should not bypass geometric perception or safety monitors. Road benchmarks need weather, lighting, missing frames, ambiguous signs, emergency vehicles, and regional rule variations.

Indoor Autonomy

Indoor VLM/VLA use cases include instruction following, pallet/dock reasoning, worker interaction, and exception handling. Benchmarks should include WMS prompt errors, occluded humans, reflective floors, motion blur, and "unknown object" handling.

Outdoor Industrial Autonomy

Yards, ports, mines, construction sites, and campuses need VLMs for semantic anomaly detection and operational context. Benchmarks should cover temporary signage, work zones, manual spotters, low-connectivity instructions, and site-specific rules.

Airside Autonomy

Airside is especially sensitive to language hallucination because operational text can be authoritative. A wrong answer about clearance, aircraft status, jet blast, or marshaller intent can cause a high-severity event. Airside VLMs should advise or constrain a planner, not directly actuate without a safety gate.

---

Airside Reliability Benchmark Design

Question Categories

Category	Example question	Expected evidence
Object grounding	"Which GSE is blocking the ego route?"	3D track, image region, map position
Aircraft state	"Is the aircraft parked, taxiing, or under pushback?"	Aircraft motion, tug connection, lights, V2X/ops state
Personnel risk	"Is any ground crew member likely to enter the ego path?"	Person track, occlusion, current task context
Rule compliance	"May the vehicle cross this hold line?"	Clearance message, map zone, default-deny rule
FOD reasoning	"Is the object on the apron likely FOD?"	Small-object detection, persistence, classification
Jet blast/intake	"Is ego inside an active engine hazard zone?"	Aircraft type, engine state, hazard polygon, thermal/V2X if available
Instruction following	"Proceed to stand B7 but avoid de-icing zone D2."	Route map, active zones, task message
Explanation	"Why should ego stop here?"	Link to object/rule/hazard that triggered stop

Corruption Matrix

Corruption	Airside examples
Visual weather	rain, fog, wet apron reflections, snow/de-icing residue
Lighting	night apron lighting, glare from terminal floodlights, low sun
Sensor	camera dropout, LiDAR partial dropout, radar ghost targets, calibration drift
Motion/transmission	frame loss, compression, stale V2X, delayed airport ops feed
Prompt/context	wrong stand ID, stale clearance, contradictory task message, missing NOTAM
Domain shift	new aircraft type, unfamiliar GSE, temporary construction layout

Evaluation Controls

• Text-only prompts must be included for every safety-relevant question.
• Wrong-context prompts should test whether the model follows invalid instructions.
• "I do not know" must be scored positively when evidence is insufficient.
• Explanations must cite observable evidence or structured inputs.
• VLA action outputs must pass an independent trajectory safety checker.
• Any model that fabricates clearance should fail the scenario, even if the route trajectory is otherwise safe.

---

Implementation Notes

Model Integration

Use four deployment modes with different acceptance criteria:

Mode	Allowed authority	Required benchmark evidence
Offline labeler	Data curation only	Accuracy and hallucination rate on withheld labels
Operator co-pilot	Human-readable advisory	Low false-negative rate for critical hazards, calibrated confidence
Planner constraint provider	Can request slow/stop/reroute through safety gate	Grounding, corruption robustness, deadline compliance
Direct VLA planner	Outputs trajectory/action	Closed-loop action validity, safety monitor compatibility, fallback behavior

Evaluation Harness

1. Freeze prompts and structured context format.
2. Run clean, corrupted, text-only, and wrong-context variants.
3. Record answer, rationale, cited objects/regions, confidence, and latency.
4. Convert VLA outputs to a standard trajectory format.
5. Run independent safety and rule checkers.
6. Report per-category results, not only an aggregate score.

Acceptance Thresholds for Airside POC

Early-stage thresholds should be conservative:

• Zero fabricated clearances in safety-hard evaluation.
• Text-only performance no better than an abstention baseline on visual questions.
• Critical hazard recall above operator-defined threshold before advisory use.
• Corruption delta explicitly bounded for each ODD condition.
• All VLA trajectories must pass independent collision, zone, and control-envelope checks.

---

Failure Modes

Failure mode	Example	Mitigation
Visual hallucination	Mentions traffic light or pedestrian that is not visible	Text-only controls and grounding labels
Language-prior leakage	Guesses common driving answer from question wording	Balanced labels and text-only score reporting
Prompt over-trust	Obeys stale "cleared to cross" text despite no valid clearance	Structured clearance schema and safety checker
Corruption brittleness	Water splash or glare changes planning answer	Corruption matrix and ODD-aware fallback
Spatial reasoning error	Misidentifies left/right side of aircraft or object distance	3D grounding and map-coordinate questions
Explanation mismatch	Says "stopping for worker" but trajectory proceeds	Explanation-action alignment check
Latency miss	Correct answer arrives after planner deadline	Deadline-aware scoring
Overconfident wrong answer	High confidence on unknown object or occluded zone	Calibration and abstention scoring
Direct-control misuse	VLA emits unsafe trajectory	Simplex safety gate and independent planner/controller checks

---

Related Repo Docs

Document	Relevance
Vision-Language Models for Airside Scene Understanding	VLM deployment patterns and airside applications
VLA Models for Driving	VLA architecture and action-head context
VLA Distillation and Scaling	Distilling large reasoners into deployable models
Spatial Foundation Models for Airport	Spatial reasoning and airport adaptation
Evaluation Benchmarks: NAVSIM and Bench2Drive	E2E planning benchmark context
Airside Autonomy Benchmark Spec	Domain-specific VLM/VLA test track
End-to-End World Model Pipeline	VLA integration as slow reasoner
Online Perception Monitoring and ODD Enforcement	Runtime gating and fallback
Simplex Safety Architecture	Authority boundaries for VLM/VLA outputs

---

Sources

• DriveBench project page
• Are VLMs Ready for Autonomous Driving?
• CODA-LM: Automated Evaluation of Large Vision-Language Models on Self-Driving Corner Cases
• CODA-LM arXiv
• DriveLM GitHub repository
• DriveLM paper
• RoboDriveVLM / RoboDriveBench
• LingoQA: Video Question Answering for Autonomous Driving
• OpenDriveVLA
• LMDrive