Letian (Zac) Chen

EDUCATION

Master of Science: Computer Science
Georgia Institute of Technology

Bachelor of Science: Computer Science and Technology
Peking University

Bachelor of Science: Psychology
Peking University

RESEARCH

Learning Interpretable Tree-based Control Policies for Autonomous Driving

• Developed interpretable, tree-based continuous-control models that allow gradient updates.
• Demonstrated the strong qualitative and quantitative performance of the proposed model in comparison with black-box neural networks in 10+ driving scenarios.
• Verified interpretability with user-studies to show the proposed model is easier and faster to interpret than neural networks and other interpretable models.

Learning from Offline Heterogeneous Demonstrations

• Analyzed real Mars rover driving data and identified heterogeneity among rover drivers.
• Proposed a novel IRL framework, DROID, to accommodate the offline learning required by the application while allowing learning from heterogeneous demonstrations via dual reward and policy distillation.
• Applied DROID on two simulated robotic control tasks and the real Mars rover path-planning problem; achieved better learning and generalization to unseen conditions in all three domains.

Fast Lifelong Adaptive Learning from Demonstrations

• Analyzed the personalization problem in lifelong learning from demonstration process where large number of heterogeneous demonstrations arrive sequentially by federation among users.
• Proposed a novel IRL framework, FLAIR, to provide efficient personalization and scalability by constructing policy mixtures with a concise set of prototypical strategy policies.
• Applied FLAIR on three virtual robotic control tasks and a real robot table-tennis task; achieved better personalization with significantly higher sample efficiency.

Learning from Suboptimal Demonstration via Self-Supervised Reward Regression

• Characterized policy performance degradation from noise injection by a sigmoid function.
• Proposed a novel IRL framework, SSRR, to learn policies that are better than suboptimal demonstrations by inferring the idealized reward function (i.e., the latent intent of the demonstrator).
• Applied algorithm on three virtual robotic tasks and a real robot table-tennis task; achieved accurate recovery of the demonstrator intention and a better-than-best-demonstration policy.

Learning from Heterogeneous Demonstrations

• Modeled humans' latent objective via shared task reward and individual strategy reward.
• Proposed a novel IRL framework, MSRD, to jointly infer task reward and strategy reward to gain a better estimation of both.
• Applied algorithm on two virtual robot control tasks and one real robot table-tennis task; achieved better learning of task reward than SOTA AIRL, extracted precise strategic rewards, and optimized versatile policies that resemble the heterogeneous demonstrations.

AWARDS

First place in Brainhack ATL 2019 Track 2

Graduate of merit in Beijing

Excellent Graduate in Peking University

Zhang Wenjin Scholarship link

Scholarship for Undergraduate Research

First Prize of National Olympiad in Informatics in Provinces Advanced Group