Research

My research goal is to enable complex robotic systems to learn complex tasks. Thus, I am interested in developing a scalable robot learning framework that can leverage prior knowledge and diverse data.

  • Skill-based learning extends the capability of robot learning to tasks with a much longer horizon via informed exploration and temporally abstracted policy learning.
  • Unsupervised skill and prior learning from large diverse data enables acquiring a suite of robust and reusable general-purpose skills and statistical prior of skills, which can lead to fast learning of a new task.
  • Robot learning in the real world requires high sample efficiency and safe exploration. This can be accelerated with the extensive use of large task-agnostic data and simulators. Another challenge here is minimizing human supervision during training.

Particularly, as the first milestone, I focus on solving a long-horizon manipulation task, Furniture Assembly, which requires many aspects of intelligent robots from structural reasoning to long-term planning to sophisticated control.

Key papers: IKEA Environment, T-STAR, SPiRL, Transition-RL, IDAPT


Progress

Long-horizon task benchmark IKEA Environment (2019)
Skill composition Transition-RL (2019), Skill-Coordination (2020), T-STAR (2021)
Learning with skills and skill prior SPiRL (2020), SkiLD (2021), SkiMo (2022)
Imitation learning SILO (2019), Goal Proximity IL (2021)
Safe exploration MoPA-RL (2020)
Simulation-to-real IDAPT (2021), MoPA-PD (2021)