We provide a sober look at the application of Multimodal Large Language Models (MLLMs) in autonomous driving, challenging common assumptions about their ability to interpret dynamic driving scenarios. Despite advances in models like GPT-4o, their performance in complex driving environments remains largely unexplored. Our experimental study assesses various MLLMs as world models using in-car camera perspectives and reveals that while these models excel at interpreting individual images, they struggle to synthesize coherent narratives across frames, leading to considerable inaccuracies in understanding (i) ego vehicle dynamics, (ii) interactions with other road actors, (iii) trajectory planning, and (iv) open-set scene reasoning. We introduce the Eval-LLM-Drive dataset and DriveSim simulator to enhance our evaluation, highlighting gaps in current MLLM capabilities and the need for improved models in dynamic real-world environments.

Embodied foundation models are increasingly performant in real-world domains such as robotics or autonomous driving. These models are often deployed in interactive or assistive settings, where it is important that these assistive models generalize to new users and new tasks. Diverse interactive data generation offers a promising avenue for providing data-efficient generalization capabilities for interactive embodied foundation models. In this paper, we investigate the generalization capabilities of a multimodal foundation model fine-tuned on diverse interactive assistance data in a synthetic domain. We explore generalization along two axes: a) assistance with unseen categories of user behavior and b) providing guidance in new configurations not encountered during training. We study a broad capability called Open-Set Corrective Assistance, in which the model needs to inspect lengthy user behavior and provide assistance through either corrective actions or language-based feedback. This task remains unsolved in prior work, which typically assumes closed corrective categories or relies on external planners, making it a challenging testbed for evaluating the limits of assistive data. To support this task, we generate synthetic assistive datasets in Overcooked and fine-tune a LLaMA-based model to evaluate generalization to novel tasks and user behaviors. Our approach provides key insights into the nature of assistive datasets required to enable open-set assistive intelligence. In particular, we show that performant models benefit from datasets that cover different aspects of assistance, including multimodal grounding, defect inference, and exposure to diverse scenarios.

Effective coaching requires instruction to be tailored to each student’s skill level. While prior work has shown the importance of adaptive feedback, there is little quantitative evidence showing how coaching language itself changes as a function of skill level - particularly in real-time embodied domains. To address this gap, we collected a naturalistic dataset of an expert coach providing verbal instructions to students of varying skill in a high-performance driving simulation. Using linguistic and interaction measures, we show that, as skill improves, coaching language shifts from concrete, instructional feedback, towards evaluative and affective language. At the same time, while overall feedback amount decreases, language diversity increases, indicating a move toward denser, more expressive communication. Together, these findings provide us with a better understanding of how a human coach modulates language as a function of learner skill and provide empirical grounding for the design of adaptive and personalized AI coaching systems.

Predicting surgical needle trajectories from endoscopic video is critical for robot-assisted suturing, enabling anticipatory planning, real-time guidance, and safer motion execution. Existing methods that directly learn motion distributions from visual observations tend to overlook the sequential dependency among adjacent motion steps. Moreover, sparse waypoint annotations often fail to provide sufficient supervision, further increasing the difficulty of supervised or imitation learning methods. To address these challenges, we formulate image-based needle trajectory prediction as a sequential decision-making problem, in which the needle tip is treated as an agent that moves step by step in pixel space. This formulation naturally captures the continuity of needle motion and enables the explicit modeling of physically plausible pixel-wise state transitions over time. From this perspective, we propose SutureFormer, a goal-conditioned offline reinforcement learning framework that leverages sparse annotations to dense reward signals via cubic spline interpolation, encouraging the policy to exploit limited expert guidance while exploring plausible future motion paths. SutureFormer encodes variable-length clips using an observation encoder to capture both local spatial cues and long-range temporal dynamics, and autoregressively predicts future waypoints through actions composed of discrete directions and continuous magnitudes. To enable stable offline policy optimization from expert demonstrations, we adopt Conservative Q-Learning with Behavioral Cloning regularization. Experiments on a new kidney wound suturing dataset containing 1,158 trajectories from 50 patients show that SutureFormer reduces Average Displacement Error by 58.6

Surgical scene understanding demands not only accurate predictions but also interpretable reasoning that surgeons can verify against clinical expertise. However, existing surgical vision-language models generate predictions without reasoning chains, and general-purpose reasoning models fail on compositional surgical tasks without domain-specific knowledge. We present Surg-R1, a surgical Vision-Language Model that addresses this gap through hierarchical reasoning trained via a four-stage pipeline. Our approach introduces three key contributions: (1) a three-level reasoning hierarchy decomposing surgical interpretation into perceptual grounding, relational understanding, and contextual reasoning; (2) the largest surgical chain-of-thought dataset with 320,000 reasoning pairs; and (3) a four-stage training pipeline progressing from supervised fine-tuning to group relative policy optimization and iterative self-improvement. Evaluation on SurgBench, comprising six public benchmarks and six multi-center external validation datasets from five institutions, demonstrates that Surg-R1 achieves the highest Arena Score (64.9

Skill atrophy, the gradual decline of human capability under AI assistance, poses a safety risk in shared-control of semi-autonomous systems, where operators may be unable to distinguish their own inputs from autonomous corrections. We propose Proximal State Nudging (PSN), a shared autonomy algorithm that jointly optimizes for skill development and task performance by nudging users toward states estimated to be most learnable. We first show that PSN outperforms existing shared autonomy baselines in balancing student improvement in unassisted reward with overall shared performance, using simulated students in the classic LunarLander environment. We then present, to the best of our knowledge, the first human subject studies of a planner incorporating learning-compatible shared autonomy: across two driving tasks in the CARLA simulator (High Performance Racing and Parallel Parking, n = 60), PSN produces up to 7x larger gains in unassisted skill than standard blended shared autonomy, while incurring 50

Drivers need to build and maintain sufficient Situational Awareness (SA) of their driving environment in order to drive safely. Therefore, the ability to estimate a driver’s SA (or its lack thereof) is critical for effective driving assistance. Traditionally, a driver’s SA has been modeled as the ability to recognize objects and spatial layouts. However, SA is layered, and encompasses not only perception and comprehension, but also projection: what might happen next. To this end, we create a novel dataset with 3,320 annotated layered SA labels from 42 participants observing a diverse set of driving videos in a supervisory control context. We extract layered SA labels directly from participants’ natural language responses to SAGAT-style probes. Our analysis of how awareness varies due to high-level features about the gaze scanpaths, object properties, and distractions, uncovers systematic patterns in the depth and dynamics of SA. We also propose a simple and intuitive baseline to predict an observer’s SA from gaze scanpaths, providing a first step towards computational modeling of layered SA. We will release the code and dataset publicly at https://layeredsituationalawareness.github.io/

Learning-based automated coaching systems for complex motor tasks such as high-performance driving remain limited in the ability to be adaptive by their reliance only on local, context-dependent reasoning, failing to account for the long-term temporal nature of student learning and the cumulative impact of repeated teacher-student interactions. In this paper, we propose an imitation learning based computational model for adaptive teaching with a dedicated temporal reasoning module that can reason over the interaction history under low-data regimes. To compensate for limited amounts of interactive training data, and based on the repetitive nature of the teaching process, the model relies on a nearest neighbor retrieval and cross attention prior, reasoning only on a narrowed-down set of semantically similar past interactions with an encoder-decoder based concurrent teaching model. We validate our approach with (i) a novel semi-synthetic closed-loop longitudinal student-teacher interaction dataset based on Waymo Open Motion Dataset and (ii) a small-scale real-world naturalistic simulator race coaching dataset. Our results reveal the consistent advantage of our adaptive teaching model with the nearest neighbor retrieval and cross-attention prior over a non-adaptive baseline as well as a suite of adaptive models that differ in their choice of priors and temporal fusion mechanisms.

Understanding human skill is important for AI systems that collaborate with, coach, or assist people. Unlike typical latent variable estimation problems which rely on single observations, skill is a persistent, compositional, and behaviorally grounded construct that must be inferred from patterns over time. We introduce Skill Abstraction with Interpretable Latents (SAIL), a method for modeling human skill as an interpretable, multi-dimensional construct inferred from naturalistic behavior. Our approach produces a skill embedding that is robust to transient performance fluctuations and learns a transferable representation of human subskills. Furthermore, SAIL supports skill-informed behavior prediction that generalizes across a variety of in-domain contexts. We represent each individual with a persistent skill embedding that controls a blend between expert and novice bases and is trained using counterfactual subskill swaps for disentanglement. This design encourages representations that are both robust to performance variation and structured for interpretability. We demonstrate across racing and baseball that SAIL achieves strong predictive performance and consistently improves behaviorally grounded disentanglement over the evaluated baselines, while also improving downstream AI coaching performance.

We propose a human-centered safety filter (HCSF) for shared autonomy that significantly enhances system safety without compromising human agency. Our HCSF is built on a neural safety value function, which we first learn scalably through black-box interactions and then use at deployment to enforce a novel state-action control barrier function (Q-CBF) safety constraint. Since this Q-CBF safety filter does not require any knowledge of the system dynamics for both synthesis and runtime safety monitoring and intervention, our method applies readily to complex, black-box shared autonomy systems. Notably, our HCSF's CBF-based interventions modify the human's actions minimally and smoothly, avoiding the abrupt, last-moment corrections delivered by many conventional safety filters. We validate our approach in a comprehensive in-person user study using Assetto Corsa-a high-fidelity car racing simulator with black-box dynamics-to assess robustness in "driving on the edge" scenarios. We compare both trajectory data and drivers' perceptions of our HCSF assistance against unassisted driving and a conventional safety filter. Experimental results show that 1) compared to having no assistance, our HCSF improves both safety and user satisfaction without compromising human agency or comfort, and 2) relative to a conventional safety filter, our proposed HCSF boosts human agency, comfort, and satisfaction while maintaining robustness.

Advances in artificial intelligence (AI) for surgical quality assessment promise to democratize access to expertise, with applications in training, guidance, and accreditation. This study presents the SAGES Critical View of Safety (CVS) Challenge, the first AI competition organized by a surgical society, using the CVS in laparoscopic cholecystectomy, a universally recommended yet inconsistently performed safety step, as an exemplar of surgical quality assessment. A global collaboration across 54 institutions in 24 countries engaged hundreds of clinicians and engineers to curate 1,000 videos annotated by 20 surgical experts according to a consensus-validated protocol. The challenge addressed key barriers to real-world deployment in surgery, including achieving high performance, capturing uncertainty in subjective assessment, and ensuring robustness to clinical variability. To enable this scale of effort, we developed EndoGlacier, a framework for managing large, heterogeneous surgical video and multi-annotator workflows. Thirteen international teams participated, achieving up to a 17

Field sobriety tests and breathalyzers are commonly used to prevent alcohol-impaired driving, but are expensive and time-consuming to administer. We propose a set of sobriety tests which, in contrast, can feasibly be automated and deployed to modern vehicles equipped with a driver monitoring camera. Our tests are inspired by research on the physiological effects of alcohol, with particular focus on eye movements and gaze behavior. We run an exploratory in-lab study with N=50 subjects (20 alcohol-impaired, 30 control), and train a variety of models to detect alcohol impairment. We find that, using only 10 seconds of observations of the driver, one of the four proposed tests performs comparably to existing non-breathalyzer field sobriety tests. We make our code and data available to support further research efforts to combat alcohol-impaired driving: https://toyotaresearchinstitute.github.io/IV25-beyond-breathalysers/.

Curated datasets are essential for training and evaluating AI approaches, but are often lacking in domains where language and physical action are deeply intertwined. In particular, few datasets capture how people acquire embodied skills through verbal instruction over time. To address this gap, we introduce SimCoachCorpus: a unique dataset of race car simulator driving that allows for the investigation of rich interactive phenomena during guided and unguided motor skill acquisition. In this dataset, 29 humans were asked to drive in a simulator around a race track for approximately ninety minutes. Fifteen participants were given personalized one-on-one instruction from a professional performance driving coach, and 14 participants drove without coaching. includes embodied features such as vehicle state and inputs, map (track boundaries and raceline), and cone landmarks. These are synchronized with concurrent verbal coaching from a professional coach and additional feedback at the end of each lap. We further provide annotations of coaching categories for each concurrent feedback utterance, ratings on students' compliance with coaching advice, and self-reported cognitive load and emotional state of participants (gathered from surveys during the study). The dataset includes over 20,000 concurrent feedback utterances, over 400 terminal feedback utterances, and over 40 hours of vehicle driving data. Our naturalistic dataset can be used for investigating motor learning dynamics, exploring linguistic phenomena, and training computational models of teaching. We demonstrate applications of this dataset for in-context learning, imitation learning, and topic modeling. The dataset introduced in this work will be released publicly upon publication of the peer-reviewed version of this paper. Researchers interested in early access may register at https://tinyurl.com/SimCoachCorpusForm.

Simulation plays a key role in scaling robot learning and validating policies, but constructing simulations remains labor-intensive. In this paper, we introduce ReGen, a generative simulation framework that automates this process using inverse design. Given an agent's behavior (such as a motion trajectory or objective function) and its textual description, we infer the underlying scenarios and environments that could have caused the behavior.Our approach leverages large language models to construct and expand a graph that captures cause-and-effect relationships and relevant entities with properties in the environment, which is then processed to configure a robot simulation environment. Our approach supports (i) augmenting simulations based on ego-agent behaviors, (ii) controllable, counterfactual scenario generation, (iii) reasoning about agent cognition and mental states, and (iv) reasoning with distinct sensing modalities, such as braking due to faulty GPS signals. We demonstrate our method in autonomous driving and robot manipulation tasks, generating more diverse, complex simulated environments compared to existing simulations with high success rates, and enabling controllable generation for corner cases. This approach enhances the validation of robot policies and supports data or simulation augmentation, advancing scalable robot learning for improved generalization and robustness.

People's goal-directed behaviors are influenced by their cognitive biases, and autonomous systems that interact with people should be aware of this. For example, people's attention to objects in their environment will be biased in a way that systematically affects how they perform everyday tasks such as driving to work. Here, building on recent work in computational cognitive science, we formally articulate the \textit{attention-aware inverse planning problem}, in which the goal is to estimate a person's attentional biases from their actions. We demonstrate how attention-aware inverse planning systematically differs from standard inverse reinforcement learning and how cognitive biases can be inferred from behavior. Finally, we present an approach to attention-aware inverse planning that combines deep reinforcement learning with computational cognitive modeling. We use this approach to infer the attentional strategies of RL agents in real-life driving scenarios selected from the Waymo Open Dataset, demonstrating the scalability of estimating cognitive biases with attention-aware inverse planning.

Learning motor skills for sports or performance driving is often done with professional instruction from expert human teachers, whose availability is limited. Our goal is to enable automated teaching via a learned model that interacts with the student similar to a human teacher. However, training such automated teaching systems is limited by the availability of highquality annotated datasets of expert teacher and student interactions as they are difficult to collect at scale. To address this data scarcity problem, we propose an approach for training a coaching system for complex motor tasks such as high performance driving via a Multi-Task Imitation Learning (MTIL) paradigm. MTIL allows our model to learn robust representations by utilizing self-supervised training signals from more readily available non-interactive datasets of humans performing the task of interest. We validate our approach with (1) a semi-synthetic dataset created from real human driving trajectories, (2) a professional track driving instruction dataset, (3) a track-racing driving simulator human-subject study, and (4) a system demonstration on an instrumented car at a race track. Our experiments show that the right set of auxiliary machine learning tasks improves prediction of teaching instructions. Moreover, in the human subjects study, students exposed to the instructions from our teaching system improve their ability to stay within track limits, and show favorable perception of the model's interaction with them, in terms of usefulness and satisfaction.

Learning in domains involving complex motor skills, such as performance driving, often requires feedback that is timely, personalized, and actionable. Yet many drivers rely on video and telemetry data to review their performance without guidance. We explore how conversational AI can support post-drive reflection by integrating LLM-generated coaching into an interactive review interface. In an exploratory within-subjects simulator study (n=16), participants completed laps under two conditions: one with video and data visualizations alone, and another with the same tools augmented with a conversational interface that provided verbal feedback after each lap. Conversational feedback supported short-term improvements in lap time, average speed, and steering control, and was rated as more useful and satisfying—though it also elicited slightly higher nervousness. These results suggest that conversational AI can make post-drive feedback more interpretable and actionable, particularly for drivers reviewing performance data in high-skill contexts like performance driving.

Real-time video understanding is critical to guide procedures in minimally invasive surgery (MIS). However, supervised learning approaches require large, annotated datasets that are scarce due to annotation efforts that are prohibitive, e.g., in medical fields. Although self-supervision methods can address such limitations, current self-supervised methods often fail to capture structural and physical information in a form that generalizes across tasks. We propose Compress-to-Explore (C2E), a novel self-supervised framework that leverages Kolmogorov complexity to learn compact, informative representations from surgical videos. C2E uses entropy-maximizing decoders to compress images while preserving clinically relevant details, improving encoder performance without labeled data. Trained on large-scale unlabeled surgical datasets, C2E demonstrates strong generalization across a variety of surgical ML tasks, such as workflow classification, tool-tissue interaction classification, segmentation, and diagnosis tasks, providing improved performance as a surgical visual foundation model. As we further show in the paper, the model's internal compact representation better disentangles features from different structural parts of images. The resulting performance improvements highlight the yet untapped potential of self-supervised learning to enhance surgical AI and improve outcomes in MIS.

In time-critical settings such as assistive driving, assistants often rely on alerts or haptic signals to prompt rapid human attention, but these cues usually leave humans to interpret situations and decide responses independently, introducing potential delays or ambiguity in meaning. Language-based assistive systems can instead provide instructions backed by context, offering more informative guidance. However, current approaches (e.g., social assistive robots) largely prioritize content generation while overlooking critical timing factors such as verbal conveyance duration, human comprehension delays, and subsequent follow-through duration. These timing considerations are crucial in time-critical settings, where even minor delays can substantially affect outcomes. We aim to study this inherent trade-off between timeliness and informativeness by framing the challenge as a sequential decision-making problem using an augmented-state Markov Decision Process. We design a framework combining reinforcement learning and a generated offline taxonomy dataset, where we balance the trade-off while enabling a scalable taxonomy dataset generation pipeline. Empirical evaluation with synthetic humans shows our framework improves success rates by over 40

The deployment of autonomous vehicles controlled by machine learning techniques requires extensive testing in diverse real-world environments, robust handling of edge cases and out-of-distribution scenarios, and comprehensive safety validation to ensure that these systems can navigate safely and effectively under unpredictable conditions. Addressing Out-OfDistribution (OOD) driving scenarios is essential for enhancing safety, as OOD scenarios help validate the reliability of the models within the vehicle's autonomy stack. However, generating OOD scenarios is challenging due to their long-tailed distribution and rarity in urban driving datasets. Recently, Large Language Models (LLMs) have shown promise in autonomous driving, particularly for their zero-shot generalization and common-sense reasoning capabilities. In this paper, we leverage these LLM strengths to introduce a framework for generating diverse OOD driving scenarios. Our approach uses LLMs to construct a branching tree, where each branch represents a unique OOD scenario. These scenarios are then simulated in the CARLA simulator using an automated framework that aligns scene augmentation with the corresponding textual descriptions. We evaluate our framework through extensive simulations, and assess its performance via a diversity metric that measures the richness of the scenarios. Additionally, we introduce a new "OOD-ness" metric, which quantifies how much the generated scenarios deviate from typical urban driving conditions. Furthermore, we explore the capacity of modern Vision-Language Models (VLMs) to interpret and safely navigate through the simulated OOD scenarios. Our findings offer valuable insights into the reliability of language models in addressing OOD scenarios within the context of urban driving.