World's first reinforcement learning-based transition control of a triple inverted pendulum
Vložit
- čas přidán 9. 11. 2023
- This video shows the world's first experimental results of the transition control of a triple inverted pendulum, in which the control law was developed using a Sim-to-Real reinforcement learning approach. The control law can make every transition between different equilibrium points possible. Furthermore, it demonstrates exceptional recovery performance, a feature lacking in classical transition control methods. Control was implemented using LW-RCP02 (Light-weight Rapid Control Prototyping 02), which was developed by Embedded Control Lab, and Simulink. The sampling time is 1 ms.
- Transition control of a triple inverted pendulum based on classical optimal control : • World's first video of... .
- Reinforcement learning-based transition control of a double inverted pendulum : • Reinforcement learning...
* The triple inverted pendulum and LW-RCP02 can be purchased from Sungjin Techwin. For purchase inquiries, please contact sales@switch-vr.com.
이 영상은 세계최초로 3단 도립진자에 대한 강화학습 기반 천이제어(transition control)의 실험결과를 보여준다. 제어법칙은 Sim-to-real 방식의 강화학습을 이용하여 개발되었다. 강화학습으로 구한 제어 법칙은 서로 다른 균형점간의 천이를 가능하게 한다. 더불어 매우 뛰어난 recovery 성능을 가지고 있는데 이에 비해 전통적인 천이제어 기법들은 recovery 성능을 가지고 있지 못하다. 제어기는 Embedded Control Lab에서 개발한 LW-RCP02 (Light-weight Rapid Control Prototyping 02)와 Simulink를 이용하여 구현하였다. Sampling time은 1 ms이다.
- 최적제어를 이용한 3단도립진자 천이제어 영상 : • World's first video of...
- 강화학습 기반의 2단 도립진자 천이제어 영상 : • Reinforcement learning...
* 3단도립진자와 LW-RCP02는 성진테크윈으로부터 구입할 수 있습니다. 구매문의는 sales@switch-vr.com로 해주세요.
Incredible! It moves our perceptions of limits in control
6 > 7 "the hay maker" 😂
seeing this gives me goosebumps, well done
i'm in awe
This is amazing. Congratulations and thanks for sharing. Well done.👍
So cool. Fascinating.
Simply super...speachless
Great system !
볼 때마다 신기하네 이번엔 강화 학습이야
This is really cool but would be much cooler if it didn't use simulink.
May I ask what the clock speed of the controller that you are using?
Is there a relationship between the length of the track and the overall length of the pendulum?
If all the pendulum elements were exactly the same length and mass (or reversed), would this become unsolvable?
any infos on the control algorithmn?
Is there encoders in the pendulum joints to get position information?
How did you mount the wires to the second and third encoders so they could rotate without getting caught?
We used slip rings.
If you watched this and don't get it-
Imagine balancing a pencil on the tip of your finger. And now add another pencil to the tip of the first pencil. And now add... Get it?