Safe Driving Across Domains

Details

• Creators:
  Zhang, Peter ; Tang, Pingbo
• Corporate Creators:
  Carnegie Mellon University ; Carnegie Mellon University. Traffic21 Institute. Safety21 University Transportation Center (UTC)
• Corporate Contributors:
  United States. Department of Transportation. University Transportation Centers (UTC) Program ; United States. Department of Transportation. Office of the Assistant Secretary for Research and Technology ; United States. Department of Transportation. Federal Highway Administration
• Subject/TRT Terms:
  • Autonomous vehicles
  • Motion
  • Trajectory control
  • Design
  • Machine learning
  • Software
  • Traffic safety
  • Risk analysis
  • Benchmarks
  • Automated vehicle control
  More +
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States ; Boston Metropolitan Area ; Pittsburgh (Pennsylvania) ; Japan ; Singapore
• Edition:
  Final Report (July 1, 2025 - July 31, 2026)
• Corporate Publisher:
  Carnegie Mellon University. Traffic21 Institute. Safety21 University Transportation Center (UTC)
• Description:
  Self-driving systems are designed and tested for specific operating conditions, called operational design domains. Moving a system to new roads, new cities, or new traffic rules normally requires retraining it on large amounts of local data, which is slow, expensive, and hard to audit. This project developed and tested a different approach. Modern self-driving planners generate several candidate paths at every moment, and the vehicle executes one of them. The project's method, called CAMP, leaves the planner untouched and improves only the final choice among the candidate paths. It scores each candidate on explicit, readable measures of safety, comfort, and progress, learns how to weigh those measures by studying recorded human driving, and adjusts the weighting to the situation at hand. In tests on 66,843 driving scenes from a public benchmark, CAMP reduced the share of scenes with a safety violation from 82.0 percent to 60.2 percent, the best result possible with the given candidates. In tests across Boston, Pittsburgh, and Singapore, most safety benefits carried over to a region the system had never seen. In 480 full driving simulations in the open-source Autoware platform, vehicles using CAMP completed every run without a collision. The project's software is publicly available as open-source code at https://github.com/Fake-fate11/camp-core. The results show that improving the final choice is a practical, low-cost, and auditable way to extend automated driving safely to new places.
• Format:
  PDF
• Funding:
  69A3552344811
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/93543/dot_93543_DS1.pdf
• File Type:
  [PDF - 1.17 MB]
• Collection(s):
  Autonomous Vehicles ; University Transportation Centers
• Main Document Checksum:
  urn:sha-512:2293a9b621eac4e4ee288c38efe7f2998097dc3f92c2591db7d906e6e36194aa3f30a1ba20fc970fa0e0a7bdac826081fc4cb8d8b21abb7d90b421c9a30e4fa8