Ms. Toru Lin

PhD Student

University of California, Berkeley

My research aims to enhance dexterity of robotic systems using end-to-end learning methods.

 

Specifically, I focus on the following topics:

 

Learning Methods:
● Reinforcement Learning:

• Training dexterous manipulation policies through deep reinforcement learning algorithms such as Proximal Policy Optimization.
• Improving the sample efficiency of reinforcement learning algorithms by encouraging unsupervised exploration via a general and flexible framework to extract intrinsic rewards.

● Imitation Learning:

• Collecting human demonstrations through a scalable and low-cost teleoperation system designed for bimanual arms with dexterous end-effectors.
• Applying imitation learning methods, such as Diffusion Policies, to equip robots with dexterous manipulation skills by learning from human demonstrations.

● Sim-to-Real Transfer:

• Addressing the challenge of zero-shot transferring skills learned in simulated environments to real-world applications.
• Employing techniques like domain randomization and enhanced physical modeling to reduce the gap between simulation and reality.

● Generative Models:

• Utilizing generative models to capture and replicate complex and multi-modal real-world data distributions, enhancing the learning process.

 

Hardware Systems:
● Bimanual Arms with Dexterous Hands

• Focusing on the development and control of robotic arms equipped with dexterous hands for complex manipulation tasks.

● Humanoid Robots

• Working on robots with human-like morphology to perform a wide range of tasks requiring dexterity and adaptability.

My works integrate advanced learning techniques with complex hardware systems to advance the field of dexterous manipulation and humanoids.