LiDAR is the safest perception sensor for autonomous vehicles being invariant to lighting condition. When dealing with LiDAR data, perception stacks mostly uses projection methods to identify vehicles. While this method is fast, it is not completely robust. We propose a pipeline to do point wise and instance segmentation on point clouds using only LiDAR data. Point wise segmentation and instance segmentation will provide the information about each point's class and which object the point belongs to. Point wise methods utilize the LiDAR data efficiently allowing for more accurate predictions of the target position and pose in the frame. We believe point wise methods will make a more robust and camera independent pipeline for detecting vehicles and pedestrians for an autonomous vehicle.
Instead of relying on drivers to communicate their intentions, which they often will not or cannot do, we take the opposite perspective; the navigation system itself should learn to predict the intentions and future behaviors of the driver based on past observations and the current situation. One approach to route prediction is to assume the driver will also try to take this most expedient route. Two important applications of this problem are route recommendation, where a driver requests a desirable route connecting two specified points, and unanticipated hazard warning, where an application can predict the driver will encounter some hazard he is unaware of and provide a warning beforehand so that the hazard can be avoided. We explore the problem of Imitation learning with two popular approaches: Maximum Entropy Inverse Reinforcement Learning and Generative Adversarial Imitation Learning and compare the advantages and drawbacks of each.
Generating jokes is a challenging and understudied task of Natural Language Processing. A computer that intends and succeeds to generate jokes could be deemed artificially intelligent. We present a couple of novel approaches to joke generation, such as using SeqGAN and a language model. We implement a variety of word-level models to tackle parts of the joke-generation problem, namely text generation and joke classification. Ideally, merging these steps will allow for a model to write joke candidates, that are then pruned by a well-trained classifier. We train these models from a corpus of 231,657 userwritten jokes scraped from reddit.com1
Collective perception for best-of-N problems in swarm robotics involves choosing the best option out of N alternatives in a decentralized manner. We aim to investigate the effects of lying and faulty individuals on collective perception using Bayesian inference. We will focus on recognizing a floor pattern made up of black and white, using a swarm of robots. Experimentation will be done on the number of lying individuals, decision making strategies of the agents and the inter-relation between them
Timeouts are necessary for network probes and outage detection. Timeouts that over- or underestimate the necessary RTT of a system can lead to delays or false negatives in outage detection systems when a host is merely connected to a high-latency, congested network. There have been several studies that have been conducted in order to determine appropriate timeout times for protocols and networking tools, such as the paper "Timeouts: Beware Surprisingly High Delay". We have recreated several experiments originally conducted by Padmanabhan et al. in their paper, and performed various experiments to determine what causes latencies.
One of the challenges faced by Autonomous Aerial Vehicles is reliable navigation through urban environments. Factors like reduction in precision of Global Positioning System (GPS), narrow spaces and dynamically moving obstacles make the path planning of an aerial robot a complicated task. One of the skills required for the agent to effectively navigate through such an environment is to develop an ability to avoid collisions using information from onboard depth sensors. In this paper, we propose Reinforcement Learning of a virtual quadcopter robot agent equipped with a Depth Camera to navigate through a simulated urban environment.
The performance of a reinforcement learning agent depends on its exploration by interacting with the environment. A popular approach to increase the exploration is to implement Asynchronous advantage actor critic algorithm, but it requires a multi-core high-performance CPU which can multi-thread environment processes. We propose a simple scalable distributed framework to implement the Asynchronous advantage actorcritic reinforcement algorithm on multiple single or dual core systems. This framework was experimented on the open source car racing simulator-TORCS. Based on the results obtained, we conclude that a distributed approach with a high number of agents gives more exploration to the agent and reduces the training time than a single agent running on a single system.