Safety evidence for neural ADAS that engineers can audit.
FieldSpace is a deterministic safety observer and validation layer for autonomy. It runs beside existing ADAS stacks, produces repeatable go / slow / stop evidence, and does not require a fleet-scale neural training loop to start generating useful safety signals.
Deterministic safety evidence for neural ADAS systems.
The current benchmark package shows FieldSpace running on public driving logs, Waymo observer scenarios, official nuPlan closed-loop simulations, and the first shared nuPlan neural-baseline smoke scenario against UrbanDriver and PlanCNN.
Neural-only validation is hard to audit.
Learned driving models can be powerful, but they leave safety teams with a hard question: why did this scene pass, fail, or change behavior after a model update? FieldSpace adds a deterministic layer around that question.
Every edge case needs a repeatable record.
Safety teams need to replay the same scene and inspect the same intermediate signals, not only trust a model score.
A good action still needs a reason.
A neural stack may choose the right maneuver, but reviewers still need traceable risk, constraints, and timing behind the decision.
Edge-case coverage is expensive to argue.
Fleet data, labeling, retraining, and scenario curation are costly. A deterministic observer gives teams another measurement path.
Add deterministic evidence where neural systems are hardest to inspect.
FieldSpace converts scene state into an auditable risk field, then emits repeatable go / slow / stop evidence. It can run beside an existing ADAS stack before anyone asks it to control a vehicle.
One deterministic observer path.
Scene state to evidence.
Five stages, all auditable, all deterministic. Every observer output is a function of its inputs, with no hidden training state or inference variance.
Perception
Camera + YOLO-class detector + Kalman tracker → object tracks with velocity.
HD Map
Lanelet2 / OSM map, Frenet projection, route planning with lane-change cost.
Prediction
1.5 s motion horizon. Map-aware lane-following, CV/CTRV kinematic fallback.
PDE Field
Continuity + velocity + potential PDEs on 256×64 grid. 0.2 ms solve.
Evidence
Go / slow / stop output, risk trace, and optional benchmark trajectory candidate.
A repeatable fallback trace, not a black-box alert.
When perception drops, route context breaks, or collision risk rises, FieldSpace records the active trigger, risk state, and recommended fallback phase for engineering review.
What we can evaluate today. What comes next.
The first wedge is not replacing an OEM stack. It is running a deterministic safety layer against public or partner-selected scenarios and turning edge cases into reviewable evidence.
Shadow-mode safety observer
Pilot-ready for log replay, public benchmarks, and narrow edge-case validation.
- ✓Deterministic risk-field evaluation
- ✓Replayable go / slow / stop outputs
- ✓Public comma.ai, Waymo, and nuPlan evidence
- ✓Audit traces for warning and disagreement cases
- ✓No neural policy training required
- ✓CPU-friendly benchmark runtime
Partner-selected edge cases
Focused replay on scenarios that matter to the customer.
- →Cut-in and hard-braking scenes
- →Pedestrian and intersection conflicts
- →False-positive-sensitive highway replay
- →Scenario-by-scenario audit packets
- →Scaled nuPlan neural-baseline comparison
OEM safety-layer integration
Integrator-led path from shadow evidence to production review.
- ○ASIL decomposition review
- ○SOTIF evidence mapping
- ○Interface hardening with partner stack
- ○Source review under NDA
- ○Pilot-to-production commercial model
The near-term commercial value is practical: run FieldSpace beside the current stack, measure disagreements, and inspect the safety evidence before changing production behavior.
Built for the standards conversation customers already have.
FieldSpace is not claiming vehicle-level certification. We are organizing the observer, replay, and evidence package around the frameworks OEM and Tier-1 safety teams use to review ADAS and autonomy systems.
Functional safety readiness
Supplier safety plan, SEooC assumptions, traceability, verification evidence, and tool-confidence path for validation use cases.
Triggering-condition evidence
Replayable edge cases, false-positive / false-negative review, ODD assumptions, and residual-risk documentation.
Scenario-based validation
Scenario taxonomy, ODD tags, source dataset, trigger type, review status, and pass/fail metrics for replay studies.
Safety-case structure
GSN-style argument skeletons and evidence registers that can become inputs to an OEM-owned safety case.
Cybersecurity and updates
Threat analysis, SBOM, vulnerability handling, release integrity, and update-impact planning for software delivery.
Customer data readiness
Security-control mapping for hosted replay, partner log handling, access review, retention, and supplier-quality review.
Current status: alignment and gap-assessment preparation. FieldSpace does not claim ISO certification, SOC 2, TISAX, UNECE approval, or vehicle-level compliance. Formal scope depends on assessor review and OEM integration context.
Safety Suite v1 results.
Five safety-critical scenarios run in CARLA with synthetic ground truth. Every scenario: earlier hazard detection, zero false positives, all braking margins met. Real-world replay against 182 k frames of comma openpilot drive logs below.
| SCENARIO | BASELINE LEAD | FIELDSPACE LEAD | GAIN |
|---|---|---|---|
| Falling Debris | -0.50s | +0.20s | +0.70s |
| Sudden Cut-In | -0.10s | +0.20s | +0.30s |
| Occluded Pedestrian | -0.30s | +0.20s | +0.50s |
| Stopped Vehicle | -0.40s | +0.20s | +0.60s |
| Sliding Cargo | -0.80s | +0.20s | +1.00s |
182,505 frames of public drive data.
Frame-for-frame replay against comma.ai's openpilot CI route bucket — real cars, real roads, real radar and vision. FieldSpace emitted one warning event and zero false criticals across 31 segments of real driving. 85% fewer spurious alerts than the prior observer.
Three deployment paths where the economics already work.
Closed-campus validation
Warehouse yards, ports, airport ground ops. Start with replay and shadow-mode review before any control authority.
Defense autonomy review
Deterministic, auditable safety behavior is useful for unmanned systems where review boards need traceable evidence.
Tier-1 / OEM validation
Run as an independent observer against customer logs and public benchmarks, then expand only where the evidence justifies it.
Two ways to deploy.
Start in parallel with zero control authority. For bounded research programs, the same core can also produce trajectory candidates for closed-loop benchmark evaluation.
Safety Observer
Runs alongside your existing stack. Zero control authority. Ingests fused objects, outputs hazard alerts with full explainability traces.
Bounded closed-loop evaluation
Planner wrapper for simulation and controlled evaluation. Useful for nuPlan-style comparison, failure analysis, and deciding which semantics to add next.
Neural ADAS needs
an audit layer.
Deterministic, replayable, and built for technical review. Bring the scenes that matter; we will turn them into evidence your safety team can inspect.