Leaves cause major problems for Tesla autopilot (futurezone.at)

Numerous user videos show Tesla’s "Full Self-Driving" software repeatedly misclassifying autumn leaves — especially wind‑blown, low‑contrast, moving leaf clouds — as solid obstacles and triggering emergency hard braking. Human drivers easily ignore or drive through such brief leaf plumes, but Tesla’s perception stack apparently treats them as objects on the roadway, prompting abrupt stops that, while so far not causing crashes in posted clips, create clear safety and traffic flow hazards. Technically this exposes weaknesses in a camera‑only approach: monocular vision struggles with transient, low‑mass objects, depth ambiguity, and distinguishing animate vs. inert debris without supporting sensor modalities. The failure mode suggests gaps in training data, temporal filtering/object tracking, semantic segmentation and risk modeling. It also rekindles debate over sensor fusion (radar/LiDAR) versus vision‑only systems, regulatory/public trust risks for partially automated vehicles, and practical fixes such as better fall‑season datasets, simulation augmentation, improved motion models, or adding complementary sensors to reduce false positives and unsafe braking.