Projects & Datasets

OpenLKA: open dataset of lane keeping assist from market autonomous vehicles

· 2024-05

The first open dataset benchmarking production lane-keeping-assist (LKA) behavior across multiple market vehicles, with synchronized CAN, video, and perception data.

AI dashcam for post-hurricane roadway damage assessment

An AI dashcam that turns any ordinary vehicle into a roadway inspection platform — automatically detecting potholes, flooding, downed trees, and damaged signs so agencies can assess storm damage in hours instead of days.

Scale-CDA: democratizing AI-assisted cooperative driving automation for production cars

An open-hardware, open-software toolchain that retrofits ordinary production cars for generative-AI-assisted cooperative driving automation with under US $1,000 of off-the-shelf parts.

DoTPilot: an in-vehicle AI platform for transportation agency fleets

An open-source in-vehicle system that turns an agency's existing fleet into a two-way link with the traffic management center — collecting roadway condition data on the uplink and delivering agency advisories to drivers on the downlink, for both daily operations and disaster response.

TriDrive: joint driver, vehicle, and road modeling for human-centered driving

A real-time world model for the cockpit — a single automation-conditioned transition model that jointly forecasts driver kinematics, vehicle dynamics, and road demands to power more timely, appropriate driver alerts.

ADAS-TO: naturalistic dataset of human takeovers during ADAS engagement

A large-scale multimodal naturalistic dataset and empirical characterization of over 15,000 real-world driver takeover events, spanning 327 drivers and 22 vehicle brands.

BATON: bidirectional automation transition benchmark

A multimodal benchmark for bidirectional automation-transition observation in naturalistic driving — 136.6 hours of driving from 127 drivers across 380 routes, capturing both handoffs to and takeovers from automation.

DriveDNA: a naturalistic driving-style dataset and benchmark

A large-scale multimodal naturalistic driving dataset and benchmark for driving-style identification — over 4,100 drives from 465 drivers across 115 vehicle models, isolating driver-specific behavior from environment and vehicle factors.

DriveMotion: a benchmark for driver motion forecasting

A large-scale, multi-source benchmark for driver body-motion sequence modeling and forecasting — 400 hours of in-cabin skeleton motion data from 360 drivers, standardized across heterogeneous monitoring sources with privacy protections.

VLAlert: vision-language models for adaptive driver alerting

A vision-language framework and benchmark for adaptive driver alerting — deciding per second whether a driving system should stay silent, observe, or alert, across nearly 193,000 labeled one-second segments.