In the rapidly advancing and exciting field of robotics,
my research is focused on developing robot systems
that possess interpretability, robustness, agility, efficiency,
and maybe even further, with intelligence and grace.
I believe that integrating learning with control
as learning-based controllers is essential for realizing these traits.
Thererfore, I seek to explore ways to utilizing the structural insights
of system dynamics provided by control theory,
to establish an optimized basis for learning algorithms.
To go even further,
I believe that combining control and learning into a unified framework
will not only result in robust learning-based controllers but also revolutionize
people's perceptions of intelligence.
The intelligence of so called "embodied agents" in the future
might look similar to systems that perceive their environments
as inputs andgenerate corresponding outputs, similar to actions.
Moreover, principles such as stability, controllability, and observability
may become essential in analyzing how these agents
perceive environments, decide, store and retrieve memories, and take actions.
Furthermore,
I am intrigued by the idea of applying physics-based character animation techniques to the learning process of humanoid robots.
I think it would be very amusing to watch humanoid robots move and behave like animated characters from movies or video games.
|
|
Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response
Junfeng Long*, Zirui Wang*, Quanyi Li, Jiawei Gao, Liu Cao, Jiangmiao Pang†
2024 International Conference on Learning Representations, ICLR 2024
[Project Page]
[arXiv]
[Code]
[BibTeX]
We present the Hybrid Internal Model,
a method enabling the control policy to estimate environmental disturbances
by only explicitly estimating velocity and implicitly simulating the system's response.
|
|
Train Once, Get a Family: State-Adaptive Balances for Offline-to-Online Reinforcement Learning
Shenzhi Wang*, Qisen Yang*, Jiawei Gao, Matthieu Lin, Hao Chen,
Liwei Wu, Ning Jia, Shiji Song, Gao Huang†
2023 Conference on Neural Information Processing Systems, NeurIPS 2023 Spotlight.
[Project Page]
[arXiv]
[Code]
[BibTeX]
We propose FamO2O,
a simple yet effective framework that empowers existing offline-to-online RL algorithms
to determine state-adaptive improvement-constraint balances.
FamO2O utilizes a universal model to train a family of policies with different improvement/constraint intensities,
and a balance model to select a suitable policy for each state.
|
Recipent for Academic Excellence Scholarship, 2023
Recipent for Outstanding Scientific and Technological Innovation Scholarship, 2023
|
|