Machine Learning & Control Systems Lab

We developed U-DiffPlan, an uncertainty-aware prediction and planning framework for autonomous racing. By combining diffusion-based trajectory prediction, game-theoretic strategy reasoning, and chance-constrained model predictive control, U-DiffPlan enables robust overtaking and defensive maneuvers under uncertain opponent behavior. Experiments show improved racing performance and real-time decision-making in competitive racing environments.

We developed LHD, a language-conditioned diffusion planner for multi-robot systems. LHD integrates vision-language understanding with a heat-inspired diffusion process to generate collision-free trajectories while reasoning about geometric reachability. The method improves robustness to out-of-distribution scenarios and achieves higher success rates with lower planning latency than existing diffusion-based approaches.

Group symmetries provide a powerful inductive bias for reinforcement learning (RL), enabling efficient generalization across symmetric states and actions via group-invariant Markov Decision Processes (MDPs). However, real-world environments almost never realize fully group-invariant MDPs; dynamics, actuation limits, and reward design usually break symmetries, often only locally. Under group-invariant Bellman backups for such cases, local symmetry-breaking introduces errors that propagate across the entire state-action space, resulting in global value estimation errors. To address this, we introduce Partially group-Invariant MDP (PI-MDP), which selectively applies group-invariant or standard Bellman backups depending on where symmetry holds. This framework mitigates error propagation from locally broken symmetries while maintaining the benefits of equivariance, thereby enhancing sample efficiency and generalizability. Building on this framework, we present practical RL algorithms -- Partially Equivariant (PE)-DQN for discrete control and PE-SAC for continuous control -- that combine the benefits of equivariance with robustness to symmetry-breaking. Experiments across Grid-World, locomotion, and manipulation benchmarks demonstrate that PE-DQN and PE-SAC significantly outperform baselines, highlighting the importance of selective symmetry exploitation for robust and sample-efficient RL.

We developed a symmetry-aware reinforcement learning framework for steering Equivariant Diffusion Policies (EDPs) beyond demonstration data. By proving that EDPs induce a group-invariant latent-noise MDP, we enable reinforcement learning to exploit geometric symmetries directly during policy optimization. The proposed framework improves sample efficiency, stabilizes learning under limited data, and remains robust when task symmetries are only partially valid. Experiments across navigation, locomotion, and manipulation tasks demonstrate substantial gains over standard diffusion steering methods, particularly in low-data regimes.

Recent progress in deep learning and Transformers has enabled major advances in robotics through imitation learning, reinforcement learning, and LLM-based multimodal perception. Yet standard neural networks often fail to exploit the symmetries and invariances inherent in robotic environments, requiring large datasets or heavy augmentation. Equivariant neural networks address these limitations by embedding symmetry directly into their architectures, yielding greater efficiency and generalization. This tutorial survey reviews core ideas and recent developments in equivariant deep learning and control for robotics, with emphasis on SE(3)-equivariant models that capture 3D rotational and translational structure in manipulation and control. We introduce essential group theory and Lie group concepts, outline the principles of group-equivariant network design, and examine applications in imitation learning, reinforcement learning, and geometric control. We conclude with open challenges and future directions for building robust, sample-efficient, and multimodal real-world robotic systems using equivariant methods.

This work presents a fully data-driven coaching framework to help novice racing drivers improve performance safely and effectively. Traditional coaching and existing automated tools often lack personalization and clear guidance, so we cluster sequential driving behavior to generate pseudo skill labels and train a classification model to assess corner-specific performance. To provide actionable feedback, we introduce a counterfactual explanation method based on negative SHAP values that identifies which feature adjustments would raise performance to expert levels. A customized LLM translates this feedback into intuitive natural language. In simulated racing experiments, our approach accurately predicts skill levels, clusters driving strategies, and significantly improves cornering and lap times. Drivers also retain these gains even after feedback is removed, demonstrating strong potential for scalable and interpretable driver training.

The study developed a deep learning framework to predict inferior alveolar nerve (IAN) injury using CBCT and panoramic radiographs. Five architectures were trained on 902 panoramic and 1,500 CBCT images, and both contrastive learning and Bayesian optimization markedly improved model performance, raising the best F1-score to 0.923. The optimized AI system outperformed oral and maxillofacial surgeons on CBCT cross-sectional images but performed slightly worse on panoramic radiographs. Overall, the combined approach effectively enhances IAN injury prediction, with contrastive learning addressing data imbalance and Bayesian optimization improving accuracy, while expert interpretation remains valuable for panoramic imaging.

Several studies have linked periodontitis to systemic diseases, yet a reliable clinical tool for early screening remains unavailable. This study leverages explainable artificial intelligence (XAI) to facilitate early detection by analyzing clinical features and providing individualized risk assessments. Using 30,465 participant records from NHANES, deep learning and machine learning models were trained on 9,632 all-age and 5,601 over-50 datasets, achieving AUCs of 0.858 and 0.865, respectively, for screening and diagnosis. LIME-based explainability identified key contributing factors, with age, sex, diabetes, smoking, and tissue transglutaminase emerging as the most significant, while arthritis, sleep disorders, hypertension, cholesterol, and obesity also played a role. The findings demonstrate the potential of XAI-driven risk assessment to enhance early detection and preventive strategies for periodontitis in clinical settings.

This paper introduces a novel framework that enhances the imitation of expert strategies by incorporating a scheduled cost function into the MPC. We present an inverse model predictive control (iMPC) framework, equipped with a custom MPC formulation that adeptly integrates scheduled cost functions. By employing Gaussian process (GP) regression, our framework effectively maps the connection between trajectories and their respective scheduling costs, enabling dynamic adaptation of cost functions within MPC planning. Furthermore, we present a probabilistic modeling method, combining Bayesian optimization (BO) with GP. This method is designed to create datasets that closely mimic expert-level driving behaviors, enriching the data available for training and validating our iMPC approach. We evaluate our framework by emphasizing the goodness of fit and interpretability of the reconstructed cost functions and the resulting trajectories. We validate our framework using human-in-the-loop expert data and demonstrate the superiority of our methodology by comparing it with a tracking MPC.

We developed deep learning models tailored to the automatic segmentation of temporomandibular joint components, including the temporal bone, disc, and condyle. These models underwent rigorous training and validation utilizing a dataset of 3693 magnetic resonance images from 542 patients. Upon evaluation, our ensemble model, which combines five individual models, yielded average Dice similarity coefficients of 0.867, 0.733, 0.904, and 0.952 for the temporal bone, disc, condyle, and background class during internal testing. In the external validation, the average Dice similarity coefficients values for the temporal bone, disc, condyle, and background were 0.720, 0.604, 0.800, and 0.869, respectively. When applied in a clinical setting, these artificial intelligence-augmented tools enhanced the diagnostic accuracy of physicians, especially when discerning between temporomandibular joint anterior disc displacement and osteoarthritis.