Every month I post links to the most recent research into motorcycle safety — crash data, protective equipment, rider training, road design, all of it. Here’s what caught my eye this month. It’s been a quiet four weeks. Little brand-new peer-reviewed work cleared into the databases that I hadn’t already linked in June, so this is a short list. Two pieces are worth your time, both on the human-factors and data side rather than crash testing or protective kit. MOTOR: a multimodal dataset for two-wheeler rider behaviourPaturkar, Gangisetty and Jawahar have released MOTOR, which they describe as the first large-scale, multi-view, multimodal dataset built specifically around powered two-wheelers in dense, unstructured traffic. It runs to 1,629 sequences and more than 25 hours of video from 16 riders, with synchronised front, rear and helmet cameras, rider eye-gaze from wearable trackers, on-road audio, and telemetry from GPS, accelerometer and gyroscope. Annotations cover traffic context, rider state, 12 riding manoeuvres and legality labels. Benchmarking rider-behaviour recognition, they found that combining RGB video, gaze and telemetry consistently beat any single input. Most of the four-wheel driver-assistance advances of the last decade were built on datasets like this; there hasn’t been an equivalent for bikes, and this is an attempt to close that gap.https://arxiv.org/abs/2605.22550 Situation awareness in motorcycle riders, measured on videoWijayanto and colleagues, writing in Future Transportation, ran a laboratory study of 30 Indonesian riders — 16 aged 17 to 25 and 14 over 25 — using the Situation Awareness Global Assessment Technique, the Situation Present Assessment Method and the Motorcycle Rider Behaviour Questionnaire. Riders watched standardised 45-minute video recordings of real road segments rather than using a simulator, which removes the workload of controlling the bike and isolates perception. Overall situation awareness came out low, and lowest among the young riders. It was also lower at night than during the day. The sample is small and specific to one district, so read it as a signal rather than a settled number, but the age and time-of-day pattern lines up with what the crash statistics keep showing.https://doi.org/10.3390/futuretransp6020078 That’s it for this month. If you’ve come across safety research I’ve missed, feel free to email me.

Every month I post links to the most recent research into motorcycle safety — crash data, protective equipment, rider training, road design, all of it. Here’s what caught my eye this month — two of the four pieces are on AI-based rider assistance, a field that has lagged the equivalent work done on cars. Fresh peer-reviewed indexing was light this cycle, so a couple of these reach back a little further than the usual four weeks; I’ve flagged the publication date on each. VLM-based advanced rider assistance system for motorcycle safetyA team from Honda Research Institute and the University of Maryland built a rider assistance system that uses vision-language models to read the road ahead and score hazards pixel by pixel — pothole severity, how slippery a puddle looks, the size and depth of a surface defect — then feeds those risk maps to a planner that recommends throttle and steering inputs to keep the rider clear. Tested in the CARLA simulator, it beat the baseline method on success rate and lowered hazard exposure. A preprint dated 27 May, so not yet peer-reviewed.https://arxiv.org/abs/2605.27948 An interpretable AI pipeline that talks riders through hazardsThis one treats a large language model as a co-pilot. The pipeline reads first-person motorcycle video at one frame per second, uses a multimodal model (Pixtral) to describe the scene and YOLOv8 to pick out vehicles, pedestrians and road hazards, then a small Mistral model turns it into short, imperative safety prompts. It runs light enough for on-device use. The authors, led by Andry Rakotonirainy, aim it at young and newly licensed riders making the move from supervised training to riding alone. Early-stage work, evaluated on public point-of-view datasets. Published 13 February.https://doi.org/10.3390/vehicles8020039 What predicts injury severity on rural undivided roadsSubasish Das and colleagues at Texas State University built machine-learning models on US crash data to rank the factors driving how badly a motorcyclist is hurt on rural undivided roads. Helmet use, the first harmful event in the crash, and crash speed came out as the strongest predictors of severity. Published 1 March in Scientific Reports.https://www.nature.com/articles/s41598-026-40755-5 Single-vehicle motorcycle crashes — and the passenger effectWorking from 5,253 single-motorcycle crashes, Radovan Madlenak and co-authors compared five machine-learning models. The headline finding runs against intuition: carrying a passenger was the single most important predictor of injury outcome, but not in the direction you might guess. Roughly 48% of riders carrying a passenger came away uninjured, against none of the solo riders in the dataset, which the authors attribute to more cautious riding with someone on the back. Single-lane carriageways, the bike overturning, contaminated road surfaces and collisions with fixed objects such as trees all raised the risk. The best model, a recurrent neural network, reached 79.56% classification accuracy. Published 5 February.https://doi.org/10.3390/app16031629 That’s it for this month. If you’ve come across safety research I’ve missed, feel free to email me.