High-fidelity simulation API:
Render camera, lidar, and radar from script-driven scenarios in dynamic virtual environments or real-world scans (PD Replica)
Challenges with AV and ADAS Perception Development
Safe, Reliable Perception
Edge Cases – Unique scenarios are critical for model accuracy and safety, but rarely seen in the real-world and often unsafe or impossible to capture
Observability – Changes to a perception model can have unintended consequences that are not observable until tested in the real-world
Safety – Crashes can lead to litigation, regulatory scrutiny, and diminish brand perception
Solution – Continuously evaluate both standard and edge-case scenarios in a virtual world designed to mimic the real-world
Time and Cost
Collection – It takes months to collect, label, and QA real sensor data
Evaluation – It takes weeks to test in the real-world
Staffing a team to capture, label, curate, QA, and field-test is expensive
Solution – Generate the scenarios, test suites, and datasets you need in days, not months
Scalability across regions and hardware
New vehicle rigs and sensor types require recapturing sensor data, labeling, curation, and QA
Entering new geographic markets requires new data and overcoming regulatory hurdles
Regulation is constantly changing and evolving requiring rework to perception models
Solution – Simulate new camera configurations, enviornments, and scenarios by updating code, not recapturing data
Simulated camera, lidar, and radar data
Parallel Domain Solutions
Evaluate
- Software-in-the-loop, closed-loop, and open-loop testing integration
- Observability - Nightly regression testing to monitor perception performance changes
- Perception unit testing (edge cases, weather, lighting, regional differences)
- Near validation testing in real-world location scans (PD Replica)
Analyze
- Benchmark against pre-defined or custom scenarios and sensor rigs, covering new regulations, accident data, locations, etc.
- Explore sensor rig selection, rig configurations, model selection
Train
- Scaled data generation for building or augmenting training data sets
- Simulate in real-world location scans (PD Replica)
- Build data sets covering infinite variations of edge cases, weather, geographic differences, lighting, sensor noise, etc