LAIA dataset: 15 hours of CARLA driving with human gaze labels for explainable end-to-end AV research
abursuc · x · 2026-09-18
- Researchers released LAIA (Labelled Attention for Intelligent Automobiles), a synthetic dataset of 15+ hours of closed-loop driving in CARLA from 44 participants, with synchronized eye-tracking plus RGB under six weather conditions, semantic/instance segmentation, depth, optical flow, and CAN bus signals.
- Scenarios were crafted to evoke natural driver responses; the dataset targets attention-aware end-to-end driving models, driver behavior prediction, anomalous attention detection, and model explainability.
- Project takeaways: 87.5% of the work happened in simulation, and imitation learning proved far more efficient than center-lane estimation. The team also compares human attention with perceptual attention emerging in their end-to-end driving models. Paper is on arXiv.
Related event: LAIA Dataset: 15 Hours of Annotated Driver Gaze in CARLA Simulation(2 posts)→
More from Embodied
- Hugging Face co-founder Thom Wolf shows off iterating robot head prototype — Thom_Wolf · 2026-09-18
- Open-source desk robot Sudo launches: self-hosted AI, $0 subscription, 25 founder units — wateriscoding · 2026-09-18
- Figure runs humanoid robots in 30 rented Bay Area homes with no new training; predicted a $10T company — LinusEkenstam · 2026-09-18
- Vintage robotics video: a handful of PhDs testing modular architectures in the wild — abursuc · 2026-09-18
- Robotics startup founder: hosting events self-selects the hard-to-find ML talent — DominiqueCAPaul · 2026-09-18
- Developers embrace open-source physics simulator MuJoCo for modular robot prototyping — Obijuan_cube · 2026-09-18