Summary of Fall 2022 GRASP SFI - Zhongyu Li, University of California Berkeley

00:00:00 - 00:45:00

This video discusses Zhongyu Li's work on developing a control framework for robotic systems. The framework uses a linear system to model the robot's dynamics, and can be used to control the robot's behavior without releasing the robot's controller online. This work could be used to create more reliable and safe robots that can be used in a variety of settings.

• 00:00:00 Jungy Li discusses the challenges of locomotion for robots in human environments, and how model-based and reinforcement learning approaches can be used to overcome these challenges.
• 00:05:00 This video introduces Zhongyu Li, a UC Berkeley researcher working on developing safe, autonomous robot movement. The video covers Li's work on collision avoidance, which uses a reduced model of Supreme-loaded inverted pendulum sleep to approximate cat dynamics. This model is then used to optimize robot movement in constrained environments. Finally, the video discusses how Li and his team are working on combining control surface methods with locomotion culture to create a safer, more autonomous robot movement.
• 00:10:00 This video discusses how models can be used to control robots, but also mentions that it can be difficult to tune the models due to errors and that different robots may require different models. It also discusses how a deep neural network can be used to control a robot, and how it can be validated using a high fidelity simulator. The policy can then be deployed on the robot without any tuning, and it can be effective at mimicking the reference motion and minimizing checking errors.
• 00:15:00 The video discusses Zhongyu Li's research on how to create a policy for a robot to follow, in the event that its original policy fails. The research uses deep neural networks to learn how to follow a specific path in the real world, as well as how to react to various stimuli. This research could be used to create more robust and reliable robots, which could be used in a variety of settings, such as manufacturing or service industries.
• 00:20:00 The Zhongyu Li discusses how the robot shields in a very fast looking cake by the io policy, but you can also generalize your conversion skills on different robots. The video also discusses how the robot can plan a trajectory to shoot a soccer ball to a random target.
• 00:25:00 The Model 3 Method is a model-free approach to controlling robot arms, which is similar to the traditional model-based method. However, the controller is essentially a black box, and the effect of the checking controllers for robot arm usually is not considered. In this talk, Zhongyu Li discusses how the model 3 method can be automated, and how it can be used to bridge the model-based and model-free safety and control paradigms.
• 00:30:00 The video discusses Zhongyu Li's work on a robot actor model that can be used to predict the walking speed of two different robots. The model's linearity is surprising, as it holds for all but one policy. This suggests that the robot's locomotion can be controlled using a single low-level policy.
• 00:35:00 This video discusses how Zhongyu Li, a post-doctoral researcher at the University of California, Berkeley, has developed a model-based control framework that can apply stability guarantees to robotic systems. The framework uses a linear system to model the robot's dynamics, and can be used to control the robot's behavior without releasing the robot's controller online.
• 00:40:00 This video discusses the parameters of a linear system and how Zhongyu Li, a researcher at the University of California, Berkeley, used an online identification method to find the controller for the robot. The video also mentions the global planner and local planners used in the algorithm, and explains that the global planner does not consider dynamics. Finally, the video provides a link to a website where users can access a Matlab simulator that allows them to connect to a mall and control the robot.
• 00:45:00 The Zhongyu Li, University of California Berkeley, speaks about the Dynamics model and High Fidelity simulator. He explains that the Dynamics model is based on the robots that are manufactured, and that the High Fidelity simulator is used to model the dynamics of the robots. He goes on to say that it is an easy thing to pour the model over from Matlab, and that the vision is to simulate gazebo. He notes that there are many divisions in gazebos, and that to achieve a highly dynamic emotions in the real world, it may be better to use a more sophisticated control method, such as reinforcement learning. He ends by saying that we need to develop a safer way for the robot to operate beside us, and that a more series and development on this part is needed.