Empiric Earth
Empiric Earth provides a measured, searchable record of the physical world built from over ten billion miles of real driving data, offering predictive intelligence products for road safety and autonomous vehicle validation. Its BADAS 2.0 model predicts collision risk from a single video stream in under 3 ms on edge hardware, achieving 99.4% average precision across long-tail categories while being 91× smaller than a frontier benchmark model.
- Autonomous Vehicles
- Artificial Intelligence
- Mobility & Transportation
- Data & Analytics
- Automotive Technology
- Software Only
- Robotics
- Insurance Technology
Funding
Founders
Founder details are not available yet.
Product
Problem
Traditional approaches to road safety and autonomous vehicle validation rely on synthetic data, established rules, and reactive reporting, which fail to capture the rare, consequential states that occur in real-world driving. This leaves fleets, insurers, and AV developers blind to actual risk until after crashes happen, delaying effective intervention and measurement.
Solution
Empiric Earth provides a measured, searchable record of the physical world built from over ten billion miles of real driving data collected since 2015. The company transformed its historical archive into predictive intelligence products, including BADAS 2.0, a world model that predicts collision risk and driver trajectory from a single video stream in under 3 ms per window on edge hardware. It also offers Atlas for searching road experience data, APEX for scoring autonomous systems against real events, Risk Index for geospatial risk concentration, and CityStream for live road network changes. The platform covers 50+ countries, ingests 300M+ new miles monthly, and catalogs 60M+ edge cases, all indexed by condition, behavior, asset, and outcome.
Target Audience
Primary customers include AV and robotics program leads, OEM and ADAS engineering teams, fleet safety managers, insurance underwriting and claims teams, as well as city and state transportation officials and research institutions conducting real-world driving studies.
Features
- BADAS 2.0 model delivers continuous collision-risk scoring with attention maps, running at the edge on Jetson Thor-class hardware with no connectivity required
- Zero-shot generalization to non-vehicular platforms including forklifts, sidewalk robots, quadrupeds, and drones without retraining
- Achieves 99.4% average precision across ten long-tail categories with a 4.6% false-alarm rate while being 91× smaller than the frontier model it is benchmarked against
- Model variants include 300M (BADAS 2.0), 86M (Flash), and 22M (Flash Lite) parameters, with inference under 3 ms on NVIDIA A100 and 5.9 ms on Jetson Thor
- Atlas search tool indexes 10 billion miles of road experience filterable by condition, behavior, asset, geography, or outcome
- APEX scoring system tests AV reaction time, anticipation, avoidance, and confidence against real recorded events before live deployment
- Risk Index measures where road risk concentrates before complaints and crash reports arrive, leveraging early signal detection (2.4 s earlier than published notices)