Deep Reinforcement Learning for Tunable Agents
In this project, we implemented a Deep entropy-augmented Q-Learning for two-player zero-sum/collaborative games and applied it on a high-dimensional state space Pursue-evasion game. We successfully showed that by tuning parameters, we can control the level of performance of agents.
Understanding the Role of Indoor Navigation Using Heads-up-display in Sparse Order Picking
A Google Glass app serves as an order picking assisting tool that provides real-time indoor navigation and order picking information to the end user.
An Android app that tracks the user's location and head rotation to provide fine-grained indoor navigation for user study purpose.
OptoRoom - A 3D Interactive Design Toolkit for Ubiquitous Optical Sensing Surfaces
Provide a scalable and time-saving way to support researchers simulate highly customized optical sensors, such as different spectral responsivity and field of view.
Understand what potentials it holds if optical sensors (such as OptoSense) are deployed in large quantity to the entire room.
Provide a set of prefabs and C# scripts to help researchers quickly set up a “real-world sized” room with groups of optical sensors for simulation use (e.g. find out a reasonable layout of lights and sensors before real-world implementation).
Self-implemented Speech Recognition Systems, Own Synthetic Voice and A Complete Spoken Dialog System
Spoken Dialog System: An Amazon Alexa pizza order system based on speech interaction
Implementation of related speech recognition and synthesizing systems.
A 5-Stage Pipeline MIPS CPU.
Support AXI bus with random delay
Exception and interruption support
Hand-written Multiplier and Divider
Test on FPGA with benchmark Coremark and Dhrystone: Ranking top 10% (groups)
Self-implemented Core OS Functions
Core functions of Non-Preemptive/Preemptive Kernel
Synchronization primitives, IPC mechanism - mailbox and related system calls
Virtual memory management
File system based on FUSE