| Session | Time | Presenter | Title |
|---|---|---|---|
| I (Morning) | 10:30 AM | Juanwu Lu | Model-Agnostic Variational Inference for Uncertainty-Aware Motion Prediction |
| I (Morning) | 10:45 AM | Michael Wozniak | Actor-Critic Reinforcement Learning & Simulated Multi-Agent Docking |
| I (Morning) | 11:00 AM | Jose Daniel Hoyos | Reward-Based Collision-Free Algorithm for Trajectory Planning of Autonomous Robots |
| I (Morning) | 11:15 AM | Alexa Lytle | Towards Optimization for the Coupled Electricity and CO2 Grid |
| I (Morning) | 11:30 AM | Soham Shirish Phanse | Exploration-Exploitation Guided Drone Fleet Dispatch Strategies in Uncertain Environments |
| I (Morning) | 11:45 AM | Sean Even | A Variational Approach to Geometric Mechanics for Undulating Robot Locomotion |
| II (Afternoon) | 2:15 PM | Yunsheng Ma | Multimodal Task Alignment for BEV Perception and Captioning |
| II (Afternoon) | 2:30 PM | Xiaohao Xu | MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction |
| II (Afternoon) | 2:45 PM | Madeleine Yuh | Designing Cognitively Aware Intelligent Tutoring Systems for Psychomotor Learning |
| II (Afternoon) | 3:00 PM | Jiyong Kwon | Real-Time Acoustic Localization Using a 2D Microphone Array and MUSIC Algorithm for Drone Detection |
Real-time acoustic localization plays a vital role in drone detection, much like the way living organisms use sound cues to navigate their surroundings and detect threats. Drawing inspiration from natural sensing strategies, this project explores how a 2D microphone array can provide precise direction-of-arrival estimates by leveraging the subspace-based MUSIC algorithm. Similar to nature’s reliance on both active and passive sensing, our approach capitalizes on robust signal processing techniques to identify and track sound sources in dynamic environments. This talk will illustrate how the 2D microphone array is designed to capture and process incoming sound waves, offering high-resolution estimates of both azimuth and elevation angles. By separating the acoustic data into signal and noise subspaces, we demonstrate how the system can effectively localize drones in real time, even in the presence of background noise or other clutter. The presentation will include experimental insights from both indoor and outdoor scenarios, ranging from controlled tests with speaker-generated noise to flights of a quadcopter drone, to demonstrate the performance and adaptability of this acoustic localization strategy. Our findings highlight accurate azimuth detection, although the single-plane microphone configuration presents inherent challenges for elevation estimation. Nonetheless, experimental results confirm the feasibility of using a 2D array for robust and real-time drone localization, underscoring its promise for practical applications in diverse operational environments.
Docking is a critical maneuver that is vital to orbital applications such as on-orbit servicing and active space debris removal. The traditional spacecraft docking problem can be extended to a multi-agent problem by coordinating multiple chasers to dock with a single target. This work is built on the fundamental concepts of pose estimation, multi-agent rendezvous, and reinforcement learning. It implements a learning algorithm for simulated spacecraft docking represented by 6-DoF robotic arms in the experimental environment. The provided algorithm, Advantage Actor-Critic, is a multi-agent, model-free, gradient-free algorithm with Actor/Critic networks and duplicate target networks. The training outcomes of the algorithm are observed via reward/loss functions and final docking performance. The results include convergence to the optimal policy and stable learning of Actor/Critic networks. The optimal policy yields robustness in docking. After training is complete, every episode that purely exploits the learned policy successfully docks in the subject scenario. The algorithm invokes several simplifying assumptions that are driven by the computational complexity and learning convergence challenges of the subject problem formulation. The incremental relaxation of these assumptions is concluded as a path to a high-fidelity docking algorithm. This implementation proves that robust docking performance can be accomplished via reinforcement learning, demonstrated to be suitable to the experimental context of 6-DOF robotics. This algorithm ultimately paves the way for further innovation in docking concepts including reinforcement learning and the subsequent unification with pose estimation.
The classical Traveling Salesman Problem (TSP) has long been acknowledged as a foundational challenge in mathematical optimization and combinatorial theory, representing one of the earliest and most influential problems in the study of algorithmic efficiency and decision-making under constraints in discrete spaces. In this research, we extend the TSP to suit robotic missions. This extension involves the integration of feasible trajectory planning and control inputs required for navigation between waypoints, ensuring obstacle avoidance, maximizing cumulative rewards, and strictly adhering to constraints such as state variables, input derivatives, mission time windows, and maximum allowable distances. We address the challenge as a two-level problem, where the upper level determines the sequence of waypoints to be visited, and the lower level computes the optimal trajectory through the candidate sequence. For the first level, we propose a genetic algorithm that employs a dynamic time-warping-based crossover and mutation to explore the discrete design space. For the second level, we use clothoid curves and the differential flatness property. Constraints are enforced through a penalty method. Results demonstrate the algorithm’s ability to find the optimal waypoint sequence, satisfy constraints, avoid infeasible waypoints, and prioritize high-reward ones. Simulations and experiments with a ground vehicle, a quadrotor, and a quadruped illustrate the adaptability of the proposed approach.
Addressing climate change necessitates innovative strategies to optimize energy systems and carbon capture and storage (CCS) technologies. This study proposes a systematic optimization framework for coupled electricity and carbon sequestration networks, aiming to minimize costs, reduce emissions, and enhance energy efficiency in Southeast Asia, a region with increasing energy demands, limited infrastructure, and significant climate vulnerabilities. We developed optimization models for carbon sequestration and electricity distribution networks using MATLAB's CVX program (Convex Optimization). The electricity network models generation, transmission, and demand to optimize energy production and sharing between countries while minimizing costs. The carbon sequestration network captures and transports CO2 to geological storage sites, focusing on cost-effectiveness and efficiency across capture, transport, and injection stages. These models were validated independently and integrated to assess resource-sharing potential and inter-country agreements. Preliminary results demonstrate that integrating electricity and carbon networks can significantly enhance cost efficiency and emission reductions compared to standalone approaches. This coupled system offers a scalable solution to address regional energy challenges while aligning with global climate goals. By leveraging linear power flow equations and optimization techniques, this study highlights the potential of coordinated network management in advancing sustainable development across Southeast Asia.
Uncrewed aerial systems (UAS) fleets and larger air taxi vehicles comprise new airspace entrants that target previously unserved or underserved air mobility-based operations. A significant portion of such operations are envisioned to take place within the airspace proximate to and within major urban areas. In addition to civilian and commercial use cases, there is also military interest in the usage of UAS, particularly within a battle space awareness context. In both urban civilian and military battle space applications, the fleet-wide management of environmental uncertainty will be critical to UAS mission success: such uncertainties include, e.g., urban micro-weather, wind conditions, and hazardous airspace monitored by adversaries, among others. In this work, we propose a multi armed bandit based exploration-exploitation framework that leverages differing fleet priorities to simultaneously explore (using lower-priority missions) and exploit (using higher-priority missions) the unknown environment. Using the example application of urban wind flow fields, we show that a Thompson sampling-inspired exploration approach produces more optimal results, as measured by regret-like performance metrics.
Limbless organisms of all sizes use undulating patterns of self-deformation to locomote. Geometric mechanics, which maps deformations to motions, provides a powerful framework to formalize and investigate the theoretical properties and limitations of such modes of locomotion. However, the inherent level of abstraction poses a challenge when bridging the gap between theory or simulations and laboratory experiments. We investigate the challenges of modeling motion trajectories of an undulating robotic locomotor by comparing experiments and simulations performed with a variational integrator. Despite the extensive simplifications that the model based on a geometric variation principle entails, the simulations show good agreement on average. Notably, our approach merely requires the knowledge of the \emph{dissipation metric} -- a Riemannian metric on the configuration space, which can in practice be approximated by means closely resembling \emph{resistive force theory}.
Bird's eye view (BEV)-based 3D perception plays a crucial role in autonomous driving applications. The rise of large language models has spurred interest in BEV-based captioning to understand object behavior in the surrounding environment. However, existing approaches treat perception and captioning as separate tasks, focusing on the performance of only one of the tasks and overlooking the potential benefits of multimodal alignment. To bridge this gap between modalities, we introduce MTA, a novel multimodal task alignment framework that boosts both BEV perception and captioning. MTA consists of two key components: (1) BEV-Language Alignment (BLA), a contextual learning mechanism that aligns the BEV scene representations with ground-truth language representations, and (2) Detection-Captioning Alignment (DCA), a cross-modal prompting mechanism that aligns detection and captioning outputs. MTA integrates into state-of-the-art baselines during training, adding no extra computational complexity at runtime. Extensive experiments on the nuScenes and TOD3Cap datasets show that MTA significantly outperforms state-of-the-art baselines, achieving a 4.9% improvement in perception and a 9.2% improvement in captioning. These results underscore the effectiveness of unified alignment in reconciling BEV-based perception and captioning.
Real-time multi-agent collaboration for ego-motion estimation and high-fidelity 3D reconstruction is vital for scalable spatial intelligence. However, traditional methods produce sparse, low-detail maps, while recent dense mapping approaches struggle with high latency. To overcome these challenges, we present MAC-Ego3D, a novel framework for real-time collaborative photorealistic 3D reconstruction via Multi-Agent Gaussian Consensus. MAC-Ego3D enables agents to independently construct, align, and iteratively refine local maps using a unified Gaussian splat representation. Through Intra-Agent Gaussian Consensus, it enforces spatial coherence among neighboring Gaussian splats within an agent. For global alignment, parallelized Inter-Agent Gaussian Consensus, which asynchronously aligns and optimizes local maps by regularizing multi-agent Gaussian splats, seamlessly integrates them into a high-fidelity 3D model. Leveraging Gaussian primitives, MAC-Ego3D supports efficient RGB-D rendering, enabling rapid inter-agent Gaussian association and alignment. MAC-Ego3D bridges local precision and global coherence, delivering higher efficiency, largely reducing localization error, and improving mapping fidelity. It establishes a new SOTA on synthetic and real-world benchmarks, achieving a 15x increase in inference speed, order-of-magnitude reductions in ego-motion estimation error for partial cases, and RGB PSNR gains of 4 to 10 dB.
Intelligent Tutoring Systems (ITSs) mimic human tutors by closing the loop between learners and tutoring agents. However, developing ITSs for psychomotor learning is challenging, as assessing correctness in traditional contexts does not directly translate to evaluating correct performance in psychomotor tasks. Key challenges include creating a knowledge space of the task, personalizing the agent to the learner’s characteristics, and maintaining learner motivation. To address these issues, we propose a cognitively aware ITS for psychomotor learning. Cognitive factors such as self-confidence influence learners' self-efficacy and learning outcomes, yet their operationalization in psychomotor ITSs remains limited. Our system incorporates an automation assistance algorithm based on an optimal control policy designed to calibrate self-confidence relative to performance. This policy is trained using reinforcement learning methods with a self-confidence Markov Decision Process framework. Additionally, we introduce a learning stage classifier to quantitatively characterize novice-to-expert transitions, bridging qualitative and quantitative representations of learning stages. By combining learning stage classification, task performance metrics, and automation assistance, our system generates tailored formative feedback—positive, neutral, or negative—enhancing personalization. This approach addresses the three critical challenges of psychomotor ITS design, offering a framework for effective ITSs.
Real-time acoustic localization plays a vital role in drone detection, much like the way living organisms use sound cues to navigate their surroundings and detect threats. Drawing inspiration from natural sensing strategies, this project explores how a 2D microphone array can provide precise direction-of-arrival estimates by leveraging the subspace-based MUSIC algorithm. Similar to nature’s reliance on both active and passive sensing, our approach capitalizes on robust signal processing techniques to identify and track sound sources in dynamic environments. This talk will illustrate how the 2D microphone array is designed to capture and process incoming sound waves, offering high-resolution estimates of both azimuth and elevation angles. By separating the acoustic data into signal and noise subspaces, we demonstrate how the system can effectively localize drones in real time, even in the presence of background noise or other clutter. The presentation will include experimental insights from both indoor and outdoor scenarios, ranging from controlled tests with speaker-generated noise to flights of a quadcopter drone, to demonstrate the performance and adaptability of this acoustic localization strategy. While our results highlight accurate azimuth detection, we also reveal limitations in elevation estimation due to the single-plane microphone configuration. To address this, future work will explore adopting a 3D microphone array to improve elevation estimation. Additionally, efforts will focus on refining algorithms for noise reduction and false alarm detection, aiming to develop a more robust system for reliable drone detection under diverse operating conditions.