Research Projects

Gaussian Process Mapping

There exist many accurate and efficient methods exist that address the mapping problem but most assume that the occupancy states of different elements in the map are statistically independent. Correlation is important not only for improved accuracy but also for quantifying uncertainty in the map and for planning autonomous mapping trajectories . Recent work proposes Gaussian Process to capture covariance information and enable resolution-free occupancy estimation. The drawback of techniques based on GP regression (or classification) is that the computation complexity scales cubically with the length of the measurement history.


Our goal is to speed up the process such that online updating is possible for large scale map without compromising too much on the precision while capturing the correlation of the occupancy of map elements. We proposed two modifications on gaussian process approach.

First, in stead of gaussian process regression, we solve the classification problem by keeping the occupancy grid representation of the map and model the binary nature of occupancy measurements (why would we need such high resolution). Second, we use information filter to replace gaussian process which is easier to compute . In order to still keep the correlation information, we apply the kernel function used in gaussian process which captures the spatial correlation between points as covariance matrix for prior distribution. The object is to prove that the error between the estimates provided by our method and those provided by GP classification is negligible.

Project under Professor Nikolay Atanasov in ECE Department at UCSD

Picar Mapping / Planning

Upcoming Planning Video...

Using DQN and DPG to Play Super Mario

Using machine learning (ML) algorithms to play computer games has been widely investigated for giving directions on achieving general artificial intelligence. Reinforcement learning (RL) is a widely studied and promising ML method for implementing agents that can learn decisions, with a human-like learning behavior. Super Mario, one of the most popular games of all time, is simple enough to provide computational traceability and yet requires highly intricate strategies. In this final project, we study applying different RL methods to teach an agent to play the game Super Mario, specifically Deep Q network(Double DQN and prioritized experience replay) and policy gradient.

Deep Deterministic Policy Gradient for Robotic Arm Control

Purely physics-model based robot manipulation has been a huge struggle for years due to the insurmountable challenge of modeling everything from physics. Machine learning, as one of the most recent applications in computer science, has proven to be a solution in robot manipulation in its great advances in estimating complex systems.The objective of this project is to use reinforcement learning to approximate the physics of a falling object in simulation and teach a robot to track the object. A camera with rapid image capture capabilities will be mounted on the robot. Visual serving is applied to process images taken from the camera, and by using deep deterministic policy gradient algorithm, the robot will learn physics through training and eventually track a falling object successfully.

Sequential Effect on Human Mind Decision

Decision Making Comparison between Human and rat.

When doing the same decision, previous experiments showed that rats take more time to make a sensory decision when the visual ambiguity is greater. When cued to prioritize accuracy, rats take more time to decide and are more accurate for any given stimulus difficulty. In these respects rat decision-making resembles that reported for monkeys and humans. But among the interleaved trials within a block, for rats the probability of their response being correct is higher in the trials with later decisions. Primates, however, are widely reported to make more errors in later responses. In this project, we tried to use sequential effect to explain the difference between human decision process and rat decision process.

Project under Professor Pamela Reinagel at Biology and Cognitive Science at UCSD (Dec. 2016 ~ Oct. 2017)


I'm a paragraph. Click here to add your own text and edit me. Let your users get to know you.

Contextual Robotics Institution at UCSD

© 2018 by Scientist Personal. Proudly created with

  • LinkedIn Clean Grey
This site was designed with the
website builder. Create your website today.
Start Now