While most autonomous driving efforts reported are directed for general driving and mainly on maj... more While most autonomous driving efforts reported are directed for general driving and mainly on major roads, there are numerous applications for autonomous vehicles for last mile mobility-from person mobility and mail delivery to flexible recharging of cars in parking structures. Over the last year, we have designed vehicles for the micro-mobility challenge. Our approach was based on adoption of the open source Autoware system. The system was taken as a starting point for the design of a robust solution. Proposed requirements include a robust control design, a shift towards increased use of image data over LiDAR data, handling of a richer set of vehicles / pedestrians in a last mile scenario, and overall system characterization and evaluation. We present an overview of the overall design and the design decisions for construction of vehicles for last-mile delivery.
Traditionally, heuristic methods are used to generate candidates for large scale recommender syst... more Traditionally, heuristic methods are used to generate candidates for large scale recommender systems. Model-based candidate generation promises multiple potential advantages, primarily that we can explicitly optimize the same objective as the downstream ranking model. However, large scale model-based candidate generation approaches suffer from dataset bias problems caused by the infeasibility of obtaining representative data on very irrelevant candidates. Popular techniques to correct dataset bias, such as inverse propensity scoring, do not work well in the context of candidate generation. We first explore the dynamics of the dataset bias problem and then demonstrate how to use random sampling techniques to mitigate it. Finally, in a novel application of fine-tuning, we show performance gains when applying our candidate generation system to Twitter’s home timeline.
Author(s): Binnani, Sumit | Advisor(s): Christensen, Henrick I | Abstract: Self-driving vehicle t... more Author(s): Binnani, Sumit | Advisor(s): Christensen, Henrick I | Abstract: Self-driving vehicle technology has become a popular topic for discussion and debate in the modern day. Although LiDAR is one of the primary sensors being used by most of the groups working towards autonomous driving, it is very costly, prone to mechanical failure, and its localization capability is not scalable due to its dependency on high-definition maps. Also, the perception related benefits of a LiDAR can be achieved by using a sensor fusion of cameras and RADAR. Considering the drawbacks of the LiDARs, the availability of an alternate solution, and the recent progress of computer vision techniques in the last few years, we are proposing an architecture for vision-based autonomous driving. In this thesis, we outline building blocks for the development of this vision-based architecture, describe the functionality of these blocks, and provide a brief overview of existing studies and research to implement t...
While most autonomous driving efforts reported are directed for general driving and mainly on maj... more While most autonomous driving efforts reported are directed for general driving and mainly on major roads, there are numerous applications for autonomous vehicles for last mile mobility-from person mobility and mail delivery to flexible recharging of cars in parking structures. Over the last year, we have designed vehicles for the micro-mobility challenge. Our approach was based on adoption of the open source Autoware system. The system was taken as a starting point for the design of a robust solution. Proposed requirements include a robust control design, a shift towards increased use of image data over LiDAR data, handling of a richer set of vehicles / pedestrians in a last mile scenario, and overall system characterization and evaluation. We present an overview of the overall design and the design decisions for construction of vehicles for last-mile delivery.
Traditionally, heuristic methods are used to generate candidates for large scale recommender syst... more Traditionally, heuristic methods are used to generate candidates for large scale recommender systems. Model-based candidate generation promises multiple potential advantages, primarily that we can explicitly optimize the same objective as the downstream ranking model. However, large scale model-based candidate generation approaches suffer from dataset bias problems caused by the infeasibility of obtaining representative data on very irrelevant candidates. Popular techniques to correct dataset bias, such as inverse propensity scoring, do not work well in the context of candidate generation. We first explore the dynamics of the dataset bias problem and then demonstrate how to use random sampling techniques to mitigate it. Finally, in a novel application of fine-tuning, we show performance gains when applying our candidate generation system to Twitter’s home timeline.
Author(s): Binnani, Sumit | Advisor(s): Christensen, Henrick I | Abstract: Self-driving vehicle t... more Author(s): Binnani, Sumit | Advisor(s): Christensen, Henrick I | Abstract: Self-driving vehicle technology has become a popular topic for discussion and debate in the modern day. Although LiDAR is one of the primary sensors being used by most of the groups working towards autonomous driving, it is very costly, prone to mechanical failure, and its localization capability is not scalable due to its dependency on high-definition maps. Also, the perception related benefits of a LiDAR can be achieved by using a sensor fusion of cameras and RADAR. Considering the drawbacks of the LiDARs, the availability of an alternate solution, and the recent progress of computer vision techniques in the last few years, we are proposing an architecture for vision-based autonomous driving. In this thesis, we outline building blocks for the development of this vision-based architecture, describe the functionality of these blocks, and provide a brief overview of existing studies and research to implement t...
Uploads
Papers by Sumit Binnani