Date: September 18th, 2026

Time: 10:00 AM (ET)

Zoom: link

Location: NSH 4305

Type: PhD Thesis Defense

Who: Ingrid Navarro Anaya

Title: Towards Generalizable Motion Prediction under Distribution Shifts

Abstract:

Autonomous robots are increasingly expected to operate in dynamic, human-centered environments. To do so safely and efficiently, they must reason about how people move and interact. In domains like driving, social navigation, and aviation, a common approach is to learn models of human motion directly from recorded data and use the resulting priors to inform downstream systems like simulators and planning stacks.

Despite the growing availability of datasets, benchmarks, and modeling techniques, state-of-the-art methods remain unreliable for real-world deployment, often generalizing poorly to novel environments and rare events. Much of this stems from recorded datasets covering few environments and few safety-relevant events relative to what a deployed system will ultimately encounter. Broadly, the field has addressed this in four main ways: validating autonomy stacks on the road, collecting more data, synthesizing relevant scenarios, and adapting at test time. These are all valuable and necessary strategies, but each is bounded by risk, cost, the sim-to-real gap, or the difficulty of reliably detecting and adapting to a shift, respectively.

These strategies share the premise that the data we hold is insufficient. This dissertation argues that such data is also underexploited and thus pursues a complementary direction, asking how much actionable signal existing datasets already contain but current practice overlooks. We do so through a recurring paradigm we call scenario characterization: describing a scenario in terms of a property of interest and acting on that description downstream. We use this paradigm in three ways. The first is for guidance and abstraction, shaping what a model trains or optimizes. The second is for mining and benchmarking, determining what a model trains on and what is withheld to test it. The third is for analysis, fixing the basis on which results are studied. We apply these uses across two settings: in-distribution generalization, which draws mainly on guidance and abstraction, and generalization under distribution shift, which draws on all three and is where the main contributions concentrate.

We further argue that evidence for generalization is typically gathered only within a single domain, so a claim that holds there is rarely challenged elsewhere. This is largely because, outside autonomous driving, no motion domain offers comparable infrastructure and scale. This dissertation therefore contributes aviation as a testbed, introducing a large-scale framework and dataset for airport surface movement forecasting.

Through this framework and our experimental settings, we expose generalization failures that would otherwise have remained hidden in aggregate metrics. We also enable reading aviation and driving on a common basis, hinting at what transfers and what does not. Finally, we also provide early evidence that the signal recovered through our framework may help systems gain the risk awareness needed to handle real-world critical events.

Thesis Committee:

Jean Oh (co-chair)

Jonathan Francis (co-chair)

Sebastian Scherer

Andrea Bajcsy

Alexandre Alahi (EPFL)