FieldSpace provides a deterministic safety observer that runs alongside any neural ADAS stack, converting perception and map data into an auditable risk field and generating repeatable go/slow/stop signals. Its fully algorithmic pipeline produces bit‑identical outputs across replays, enabling safety teams to trace decisions, verify compliance with NHTSA and Euro NCAP protocols, and conduct offline or shadow‑mode validation on modest CPU hardware.
Funding
Funding not disclosed
Founders
Product
Problem
Autonomous driving teams struggle to audit neural ADAS decisions because existing validation relies on opaque, non-deterministic models that require large fleet data, extensive labeling, and costly retraining pipelines. This makes it difficult to reproduce edge‑case scenarios, trace the reasoning behind safe or unsafe maneuvers, and demonstrate compliance with safety protocols.
Solution
FieldSpace provides a deterministic safety observer that runs alongside any neural ADAS stack and converts scene state into an auditable risk field. The system processes perception inputs, HD‑map data, and short‑term predictions to solve a PDE on a 256×64 grid, producing repeatable go/slow/stop signals and a traceable risk explanation for each decision. Because the observer is purely algorithmic, its outputs are bit‑identical across replays, enabling engineers to replay the same scene, inspect intermediate signals, and verify compliance with NHTSA and Euro NCAP protocols without training a neural policy. The observer runs on modest CPU hardware, supports offline benchmark replay on public datasets (e.g., openpilot, Waymo, nuPlan), and can be deployed in shadow‑mode for live vehicle testing.
Target Audience
Primary customers are autonomous vehicle developers and safety engineering teams that need deterministic validation of neural ADAS components, including OEMs, Tier‑1 suppliers, and advanced driver‑assist system providers.
Features
- Deterministic risk‑field evaluation pipeline with five auditable stages: perception, HD‑map, prediction, PDE solve, and evidence generation
- Replayable go/slow/stop outputs with full audit trace linking decisions to object tracks, map constraints, and timing
- CPU‑friendly runtime (sub‑millisecond per frame) on platforms ranging from Raspberry Pi 5 to x86 laptops, no GPU required
- Compatibility with public driving logs and benchmark suites (openpilot, Waymo observer, nuPlan closed‑loop, CARLA) via provided dataset adapters
- Proven compliance checks against NHTSA 37 pre‑crash families and Euro NCAP AEB/FCW scenarios with 100 % pass rates in‑ODD
- Shadow‑mode deployment kit for live side‑by‑side evaluation on up to 10 vehicles, including weekly hazard dashboards and safety case documentation
- Integration support for AUTOSAR, ROS 2, and gRPC, with source escrow and volume‑tiered licensing for production use