Transferring agents trained in simulation to physical environment is one of the key technologies for efficiently training autonomous robots. Due to the gap between simulation and physical environment, it is crucial for the agent to be able to learn and adapt to the variance of image, physics, motion, feedback, etc. in the environment.
Robots often need to adapt and learn online in the physical environment, with high demands for sample efficiency due to high cost in the physical environment. Research of batch/offline learning, model-based learning, self-supervised learning, etc. are important to continuously optimize and improve agent learning efficiency.
To be able to generalize to unseen environments and new tasks, which is necessary for large-scale application of robots, agents need generalization capabilities in representation, multi-tasks learning, reasoning, and decision-making.
Agents with perception, navigation, and adaptive control capability can be used for motion control and planning of robotic arms, quadruped, and wheeled robots. Research areas include deep learning, reinforcement learning, multi-modal perception, planning, and control.
Use of reinforcement learning for combinatorial optimization, multi-agent reinforcement learning to optimize cooperation efficiency for a group of robots.
For business inquiries, please fill in the contact form and we will get in touch as soon as poosible.
11 / F, block B3, Zhongguancun No.1, Beiqing Road, Haidian District, Beijing
