Vehicle manufacturers are introducing advanced safety and driving aids by incorporating sensors like LiDAR, RADAR, cameras, etc. Further, there is an increased regulatory pressure to incorporate features like automatic emergency braking and collision warning in all vehicles. These require vehicles to have data-driven prediction systems that can accurately predict maneuvers and unexpected behaviors of surrounding vehicles. We propose ManeuverStack, a Stacked Bidirectional Long Short Term Memory (LSTM) model for maneuver prediction. This online model continuously collects surrounding vehicles’ data and predicts their maneuver in the next 0.5 to 5 seconds. The model has been trained and tested on a mix of six datasets which had different road characteristics and vehicle classes. Its maximum accuracy was 95%. It has been benchmarked against six state-of-the-art models available in the literature. The model’s accuracy, recall and F1 scores outperformed most models while being at par with models which predict fewer maneuver classes. The model also performed on par with the best benchmark in crash avoidance scenarios. The model’s high accuracy ensured 100% crash avoidance even at deceleration lower than the standard recommended value for passenger comfort. When predicting maneuvers for high-risk scenarios, the model provided higher accuracy than most of the benchmark models. Further deployment and testing of the model on an edge device revealed that the model’s response time is at least two orders of magnitude faster than the minimum required prediction window.

For the immediate future, Autonomous Vehicles (AV) will have to share the road with human-driven vehicles. However, decision errors by human drivers during maneuvers result in crashes. AV therefore have to predict human-initiated safe and unsafe maneuvers to react appropriately for the safety and comfort of its passengers. We propose ManeuverNet, a deep neural network model that uses sensor data to predict if any vehicle within its field-of-view would perform safe or unsafe lane maneuvers. The maneuver could either be into its lane, or into any other vehicle's lane in the next 0.5 to 5 seconds. The model predicts five classes of maneuvers by taking fewer features than the published state-of-the-art. It distinguishes safe lane changes from unsafe cut-in maneuvers. It was pre-trained using a stacked auto-encoder network to improve the model's performance on small datasets. The model has been tested on the NGSIM and CitySim datasets. The model predicted with 93% accuracy even with just two seconds of observation. Advance braking using the model's predictions resulted in safer headways during at least 91% of the cut-in events and passenger comfort was retained during at least 77% of such events. In addition, the model provides superior accuracy in high-risk scenarios like high speed and low headway.

When a vehicle changes lane abruptly resulting in low headway distance to the vehicle behind it, it may trigger emergency braking or it may result in a collision. Such events are called cut-ins and they cause a majority of road accidents in the U.S. Autonomous Vehicles (AVs) have to predict such events in advance and react appropriately to such human behavior. We propose Predicut, a stacking classifier model that uses sensor data of an instrumented host vehicle and predicts if any of the neighboring vehicles would cut in front of it in the next 0.5 to 5.5 seconds. We have built and tested the model using the SPMD and NGSIM (I-80, US-101) driving datasets. The model was able to predict cut-ins with a maximum F1 score of 94.3%. From the datasets, we determine that the model through its advance prediction, avoided up to 90% of emergency braking scenarios. The model is highly performant with average inference time of 26ms when hosted on a cloud based serverless inference service.

Heating, ventilation, and air-conditioning (HVAC) system’s supervisory control is crucial for energy-efficient thermal comfort in buildings. The control logic is usually specified as ‘if-then-that-else’ rules that capture the domain expertise of HVAC operators, but they often have conflicts that may lead to sub-optimal HVAC performance. We propose EImprove, a reinforcement-learning (RL) based framework that exploits these conflicts to learn a resolution policy. We evaluate EImprove through a co-simulation strategy involving EnergyPlus simulations of a real-world office setting and a formal requirement specifier. Our experiments show that EImprove learns 75% faster than a pure RL framework.

Increasing real estate and other infrastructure costs have resulted in the trend of co-working offices where users pay as they use for individual desks. Co-working offices that provide personalized comfort need to address users with potentially widely varying thermal comfort preferences. Providing personalized comfort in cabins separated by physical partitions with neighboring thermal zones or open-plan offices with a single actuator has received attention in the literature. In this article, the problem of minimizing user discomfort in open-plan co-working offices with multiple actuators while being cognizant of the energy consumed is considered. Specifically, the decision problems of assigning users to desks based on their thermal preferences and jointly controlling the multiple actuators are addressed. The non-linearities in the underlying thermodynamic constraints and the seating decision together make the problem computationally hard. A two-step heuristic that addresses these issues is presented. First, using a model that accounts for spatio-temporal thermodynamics, a one-time assignment of users to desks is performed that reduces the thermal resistance faced by the HVAC systems to provide the preferred comfort levels. Next, the setpoints are decided for all actuators to jointly minimize user discomfort by optimization and model-predictive control. Further, scalability is addressed by clustering user preferences and the associated HVAC actuators’ setpoints for the cases where a large number of actuators may be present in the room.