正在加载视频...
视频加载失败
Triple inverted pendulum in transition control A classic control problem, done in real time. This setup moves smoothly between all eight equilibrium points of a triple inverted pendulum. The system reacts every 1 millisecond, which shows how fast modern control loops can be. • Real-time control at 1 ms... show more
48 条评论
Great demonstration. I know a regular inverted pendulum is the principle behind Segway and similar tech. Where is the triple inverted pendulum used?
mainly employed in research and education to test advanced nonlinear control algorithms, with emerging applications in robotics for multi-joint stability, aerospace for attitude control, and biomedicine for prosthetic limb balancing
Hello 3 body problem
illustrating how modern algorithms tame chaotic systems
@Darren_Dawson And more than 3 ?
@IlirAliu_ @Darren_Dawson Give it to a control systems student and don't tell them it's any difficult
@IlirAliu_ @Darren_Dawson Yeah like cold fusion 🤭
@IlirAliu_ @Darren_Dawson I think you're confused a bit This isn't a "solution to 3 body problem" like some xeeters claim That's a completely different thing The pendulums are kept actively outside the chaotic realms. The video is not ~solving chaos nor ~predicting general movement
@ikario404 @IlirAliu_ I know, attire is dead, I forgot.
Wow. Impressive. I bet most people don't realize the complexity behind this. From modeling to actuation. What's the feedback system? Vision?
feedback system relies on four encoders measuring cart position and pendulum angles, combined with a time-varying Linear Quadratic controller for error correction, rather than vision-based sensing
I remeber doing this project with an RL agent during my probabilistic AI course at ETH, very fun project
And footage pics of that time?
'The system reacts every 1 millisecond, which shows how fast modern control loops can be.' Interesting point... Humans can't even perceive sub 4ms. I'm to understand this robot is operating at about .25ms to accomplish trajectory tasks.
This will be my go-to video when I'm trying to describe to a non-technical person why self-driven cars are better than human ones.
holy shit. that's superhuman
> reacts every millisecond must be written in python
1 ms sampling is slow in computer time. I worked on a microprogrammed disk controller in the early 1980s that had delay loops to wait for the next bit to be read. One bit took about 250 nanoseconds to arrive. The slow part is the positioning solenoids.
This is cool! I saw a similar animated video (Pezza's Work - YouTube) using neural nets for control.
Can you find it again?
I just sent you a DM.
This is certainly a cool project. But re "how fast modern control loops can be": I was working on systems with 1ms sampling 15 years ago, and that was on a legacy product even then. The only area where I've met people treating 1ms as fast is web.
@DanielleFong @asingleoat
sorry but 1ms is pretty slow. try usecs or even nsecs for tight loops.
Impressive. When I was studying, the single inverted pendulum was cutting edge control theory. Now kids have access to all sorts of balance balls and wheels and theory has moved on significantly.
Incredible.
Humans don't stand a chance😵
Indeed. But a more impressive inverted pendulum example is Astra’s ill fated launch 3.2 with the rocket translating sideways, stably.
Is this learnt (RL etc)? Or is it 'solved' where there's an answer/formula?
Very impressive transitions. Millisecond operations are slow though, I wonder what that comment refers to?
are you using gym to train the model which controls this or am i misreading the engine behind all this
Video that I found and thought it would be interesting to share
the comments in the yt video are perfect. describing physics as 'evil magic' is sorta lyrical from the right perspective. also surprised this is 2 years old.
I like how it thinks it needs to do a transition swing to get from the second lowest to the lowest state 😆👍
Matlab über alles 🤩
I can do the last one.
It is probably way harded to do this than to control the landing of those rocket first stages.
OK guys, we've solved chaos now, and we..... *looks at the state of the world* never mind.
Control engineering p0rn. Love it!
Really stunning! Is it done with an RL controller or with a model-based framework?
Wooooow!
This is the most impressive thing I've seen today, it's absolutely amazing how the robot manages to control a triple pendulum considering the difficulty of a two pendulum problem, this is insanely beautiful
Wow that is really amazing :)
really impressive
😱😱
Impressive! Is there a github repo?
Cool, funny thing I made a double pendulum simulation a few days ago. It targets embedded control like this, checkout if you want to go possibly faster than simulink generated code.
Not optimized code is detected! Last move was totally unnecessary, simply stop would have been sufficient to drop.
