Motion prediction has been an essential component of autonomous driving systems since it handles highly uncertain and complex scenarios involving moving agents of different types. In this paper, we propose a Multi-Granular TRansformer (MGTR) framework, an encoder-decoder network that exploits context features in different granularities for different kinds of traffic agents. To further enhance MGTR’s capabilities, we leverage LiDAR point cloud data by incorporating LiDAR semantic features from an off-the-shelf LiDAR feature extractor. We evaluate MGTR on Waymo Open Dataset motion prediction benchmark and show that the proposed method achieved state-of-the-art performance, ranking 1st on its leaderboard1.

Despite decades long establishment of effective tutoring principles, no adaptive tutoring system has been developed and open-sourced to the research community. The absence of such a system inhibits researchers from replicating adaptive learning studies and extending and experimenting with various tutoring system design directions. For this reason, adaptive learning research is primarily conducted on a small number of proprietary platforms. In this work, we aim to democratize adaptive learning research with the introduction of the first open-source adaptive tutoring system based on Intelligent Tutoring System principles. The system, we call Open Adaptive Tutor (OATutor), has been iteratively developed over three years with field trials in classrooms drawing feedback from students, teachers, and researchers. The MIT-licensed source code includes three creative commons (CC BY) textbooks worth of algebra problems, with tutoring supports authored by the OATutor project. Knowledge Tracing, an A/B testing framework, and LTI support are included.

In this paper we propose an unsupervised learning framework to predict risky driving at intersections in a connected vehicle environment. The proposed framework uses time series k-means to categorize multi-dimensional time series trajectories into several context-aware driving patterns. Dynamic time warping (DTW) is implemented within the time series k-means algorithm for measuring the similarity between trajectories. DTW is adopted to make the framework robust to temporal distortions and missing data points. We train an isolation forest (iForest) model on the trajectory dataset to identify anomalous trajectories, and apply this model to clusters to provide Risky Driving Prediction (RDP) scores for each driving pattern. We provide a real-time online assessment approach to predict the risk score of driving trajectories that travel toward a signalized intersection. We use real-world connected vehicle trajectories collected by a road-side unit (RSU) in Ann Arbor, Michigan to implement our framework. We use several quantitative measures as well as illustrations to validate our model. We further discuss how the RDP framework can be used to develop network-level and individual vehicle-level insurance and safety focused applications.

Automated vehicles are envisioned to be an integral part of the next generation of transportation systems. Whether it is striving for full autonomy or incorporating more advanced driver assistance systems, high-accuracy vehicle localization is essential for automated vehicles to navigate the transportation network safely. In this paper, we propose a reinforcement learning framework to increase GPS localization accuracy. The framework does not make rigid assumptions on the GPS device hardware parameters or motion models, nor does it require infrastructure-based reference locations. The proposed reinforcement learning model learns an optimal strategy to make “corrections” on raw GPS observations. The model uses an efficient confidence-based reward mechanism, which is independent of geolocation, thereby enabling the model to be generalized. We incorporate a map matching-based regularization term to reduce the variance of the reward return. The reinforcement learning model is constructed using the asynchronous advantage actor-critic (A3C) algorithm. A3C provides a parallel training protocol to train the proposed model. The asynchronous reinforcement learning strategy facilitates short training sessions and provides more robust performance. The performance of the proposed model is assessed by comparing it with an extended Kalman filter algorithm as a benchmark model. Our experiments indicate that the proposed reinforcement learning model converges fast, has less prediction variance, and can localize vehicles with 50% less error compared to the benchmark Extended Kalman Filter model.

We address the problem of optimizing recommender systems for multiple relevance objectives that are not necessarily aligned. Specifically, given a recommender system that optimizes for one aspect of relevance, semantic matching (as defined by any notion of similarity between source and target of recommendation; usually trained on CTR), we want to enhance the system with additional relevance signals that will increase the utility of the recommender system, but that may simultaneously sacrifice the quality of the semantic match. The issue is that semantic matching is only one relevance aspect of the utility function that drives the recommender system, albeit a significant aspect. In talent recommendation systems, job posters want candidates who are a good match to the job posted, but also prefer those candidates to be open to new opportunities. Recommender systems that recommend discussion groups must ensure that the groups are relevant to the users' interests, but also need to favor active groups over inactive ones. We refer to these additional relevance signals (job-seeking intent and group activity) as extraneous features, and they account for aspects of the utility function that are not captured by the semantic match (i.e. post-CTR down-stream utilities that reflect engagement: time spent reading, sharing, commenting, etc). We want to include these extraneous features into the recommendations, but we want to do so while satisfying the following requirements: 1) we do not want to drastically sacrifice the quality of the semantic match, and 2) we want to quantify exactly how the semantic match would be affected as we control the different aspects of the utility function. In this paper, we present an approach that satisfies these requirements.