Tuan-Anh Vu

Rather than assuming every modality is always available and reliable, I study modality-aware representation learning: objectives that exploit complementary cues during training but stay stable when a sensor is missing, degraded, or replaced. Robustness is measured not by average accuracy alone, but by calibration under shift, worst-group performance, recovery after sensor degradation, and how much new data or computation adaptation costs. The key idea is to teach by intervention: record the same scene with one sensor deliberately degraded, so that the model learns what each sensor's failure looks like.

• Selective and calibrated prediction — uncertainty estimation, out-of-distribution detection, and abstention criteria tied to the cost of downstream errors
• Data-efficient adaptation — self-supervised and test-time methods that learn from unlabeled streams without catastrophic drift
• Spatial-temporal foundation models — representations that jointly encode appearance, geometry, motion, and semantic context across 2D–4D data

The contribution here is not a faster inference engine but an observable AI pipeline, in which data provenance, model version, confidence, latency, and task outcome can be traced together. Streaming world models should maintain geometry, motion, semantics, and uncertainty as new observations arrive, across three scales usually studied in isolation.

• Training scale — distributed experiments and controlled data engines for multimodal pretraining, adaptation, and rare-event generation
• Representation scale — compact, online-updatable scene models with consistency checks, change detection, and memory management over long deployments
• Deployment scale — profiling-guided compression, cloud–edge partitioning, runtime monitoring, and fallback behavior under compute or sensor failure

I use demanding field applications as stress tests that reveal general AI and systems questions. Agricultural and marine platforms remain my development testbeds because they permit controlled study of occlusion, environmental shift, and rare failures before methods transfer elsewhere. The shared experiment across domains is to propagate a perturbation through the stack and measure how a sensing change alters confidence, latency, decisions, safety, and recovery.

• Agriculture — crop intelligence under occlusion and shift, longitudinal digital twins and precision mapping, and reliable edge AI for targeted spraying, scouting, biological pest control, and robotic pollination
• Manufacturing and inspection — multimodal detection of transparent, reflective, occluded, or novel parts, and dynamic reconstruction for robotic inspection and manipulation
• Transportation and logistics — persistent 3D/4D mapping, asset monitoring, and risk-aware mobile autonomy where latency, coverage, and recovery matter as much as recognition