Видео

Yes, I'm glad to help. What issues are you facing?

​@@stevenhong7099 I had trouble identifying the components ,If you can explain to me the idea using Captor ,What is the role of Captor here and how does it work? , thank you very much, sir

@@aichahadil287 sorry I don’t quite follow you. What is Captor you are referring to?

Glad you enjoyed it!🎉🎉

Yeah, even human can’t walk well on ice. But I did manage to make Spot walk on ice. The lesson learned is to reduce the speed when it’s fore legs touch the ice. This will reduce the moment and make robot being able to walk on the ice. That’s only for slow walk. I think engineers can manage to make Spot skating on the ice but do we really want to invest money on that lol.

Yeah ,i thought it would be more equal race , guess i was wrong 😂

Putting learned policy on the physical robot requires a lot of extra work. What we did here is not enough for the implementation on the physical robot. First, we want a trained policy that robot won't fall or get tripped in the simulation. With the current setup, robot still have a small chance of falling. The reward function still requires some tuning. Second, we need to train the ANYdrive characteristics. Motors have nonlinear models and we didn't include this into the simulations. ETH has a good paper on this. arxiv.org/abs/1901.08652 Third, the main idea behind this methodology is teacher-student policy. What we trained is the teacher policy, while the final policy implemented on the robot is the student policy. Teacher policy has the privileged information, which is not possible on the real life. So we want to use teacher policy to teach student policy, and implement the student policy on the physical robot. arxiv.org/abs/2010.11251 Fourth, the robot in our lab is ANYmal C, while we used ANYmal B in this video. We started with ANYmal B because it has simpler dynamics model and can train faster. However, when we migrated from B to C, the robot was behaving super weird (crawling on its knees). It means that the tuned weights for B is not suitable for C. Since the deadline for the project was approaching soon, my team decided to stick with ANYmal B within the scope of ECE545. If we want to put this policy on the real hardware, we need to retrain the ANYmal C model with new reward function.

Hello, the idea of this work is inspired by ETH RSL lab, the hub of all ANYmal robots. They came up with an idea of teacher-student policy where teacher has all the information needed to training the locomotion policy, while student only has proprioceptive sensor inputs. The term "privileged" is basically referring to the teacher policy here. ruclips.net/video/9j2a1oAHDL8/видео.html My lab historically worked on the model-based controller and reachability-based motion plan. As you can tell, we are more dynamics people and mathematics people, so we used to "despise" the reinforcement learning because of all sorts of the issues. However, the excellent results shown by ETH blew our mind, so we start to explore this area as well. They also published a new paper incorporating the perception at the beginning of this year, which makes my paper "useless" from the initiative and motivation. ruclips.net/video/zXbb6KQ0xV8/видео.html

PID is one of controls technique widely used. It’s simple and doesn’t need to model the system to control the signals. It’s a feedback control so it requires sensors to measure the information and data. There are three parts correspond to 3 letters. P is proportional control, basically you calculate the error between current state and reference state, the control input is proportional to the error. This is the driving force of control. I is integral control. You integral the errors over time. This is just a sum-up in digital system. I term is dedicated to reduce steady-state error. D is derivative control. You calculate the derivative of two consecutive sensor data, add them to the control input. This is dedicated for speed control. For example, prevent the overshoot of the control signal.

No, there is an angle sensor at the bearing

Can you see the F-22 Air Force Team sticker on the back side of my laptop, lol. Maybe the FRC hand is more visible

I think legged robots have a big potential in the future. Showing what they can do and what they can’t really help people become more familiar with this new technology. I’m very help to see the legged robot community to grow

You know how to come up with a better RUclips title 😂😂😂

Yeah, I was controlling the robot at that time, and I didn't move the robot. It ran away by itself because of the obstacle avoidance

Spot is not on open-source and only API. ANYmal is open-source after signing documents with ANYbotics. HyQ is open-source on GitHub. Digit is not open-source and we wrote our own controller.

Boston Dynamics has the best robots in the world so they don’t give a shit of sharing source codes.😅😅😅

Explore

Robotics is a very broad area, and depending on what you are looking for, the answer may vary a lot. I'm not sure what kind of materials you are looking for, but taking robotics courses is always a good way to learn. After you getting into a research lab, you spend more time reading papers and attending seminars to learn stuff instead of robotics. The field is growing really fast and I say it's exponentially growing, so basically I'm also learning new things everyday by working on a specific project. If you narrow down the scope of the materials you are looking for, I probably could give you a better answer, but the following is very generic. 1. RSS and ICRA are the two top conferences for robotics. It will be overwhelming for new people, but you will learn the-state-of-art from there. I met my current research advisor in a RSS conference. 2. RUclips has a couple of excellent videos posted by the famous research groups. For example, I watch a lot of videos from ETH/RSL, MIT CSAIL, UPenn, UCB.

​@@stevenhong7099 I have watched some autonomous vehicle oriented courses(a 1-2 hours lecture for each topic like SLAM, perception, control, and path-planning) and then read a book about ROS and after that, I just got stuck and did not know where to go, I am interested in deep learning especially fro computer vision and did some projects and willing to read about reinforcement learning and want to get hands-on experience in these topics in robotics. So it would be great if you could suggest college courses(ones that are available online) with simulation tasks/assignments to get some theoretical and practical knowledge, and sadly there are no robotics research group/lab where I am as the field itself is just new in my country Egpyt, here we focus more on app and web development as they require no "expensive" hardware . I really appreciate your help and the conferences you have suggested are really great I have searched for them on youtube.

@@mohamedel-hadidy4844 Yeah, I understand that the hardware or equipment is expensive. Only big title universities have some advantages on the funding for getting the most advanced hardware. My undergrad is a small school, so they didn't have good robotics program as well. But for simulation, computer vision, or reinforcement learning, you could work on any good laptop. 1. For deep reinforcement learning, CS285 from UC Berkeley is the best course I've ever seen. Dr. Sergey Levine is one of the best RL professors. rail.eecs.berkeley.edu/deeprlcourse-fa20/ 2. For computer vision, CS231n from Stanford is another famous course. Dr. Fei-Fei Li is one of the best professors in this fields. cs231n.stanford.edu/ 3. For SLAM and perception of drones, I took VNA2V course at UMich. It was not open-source but MIT version is open-source. ocw.mit.edu/courses/aeronautics-and-astronautics/16-485-visual-navigation-for-autonomous-vehicles-vnav-fall-2020/index.htm 4. Also, MIT Open Courseware is a good place to learn stuff. You can try finding the topics you want there. ocw.mit.edu/courses/find-by-topic/

@@stevenhong7099 I have a trustworthy gtx1060 laptop XD so I believe it will do the job, I have taken a look at the material you have sent and they are all great and I am planning on starting on them as soon as the end of my finals, so thank you very much I really appreciate your help <3

I did Mechanical Engineering major for my Bachelor and Master degree. I also have minors in electrical and computer engineering, robotics, and economics. Currently my focus is on robotics, such as quadrupedal robots.

Woah 🥰, this is exactly who i wanna be. Thanks, ❤️❤️❤️

It's step motor with optical encoder. The payload is only 500g with gripper. We only used this for developing kinematics control algorithm. It focuses more on theory than the application

Spot definitely wins without doubt

@@stevenhong7099 and you know that, because Spot wins the drag Race?

@@cyberistful Spot can literally go everywhere except ice lake, we tested that for my lab’s newest paper. However, ANYmal still couldn’t handle stairs with the package we purchased. Also it’s blinding locomotion so no perception built into it that means no automatic obstacle avoidance. There are ETH videos showing up ANYmal but only that lab has the code and also knows how to run it. We are also struggling a lot other things with ANYmal too, such as adjusting the walking speed.

@@stevenhong7099 But what I heared from Anybotics is that, they have a ready-to-use programm codes in which we can directly modify and adapt to our controller. In other words, they have an open-source community and also opensourced their codes.

@@shengzhiwang2143 I do have access to their source codes. But it’s 1 million lines of codes

Good work ! Do you share the code ?

@@oracid It's really simple PID controller and nothing fancy. The rest is just setting up the PIC/microcontroller with PWM. That's more professional embedded system coding as we (engineers) don't use libraries. For clarify, we use registers directly so the code runs faster then using library and make more profit for the company. /*********************************************/ // implement the CONTROLLER (Gc) functions // PID controller /*********************************************/ Derror = error - last_error; last_error = error; Isum = Isum + error; if (Isum > 0.0) Isum = min(Isum, MAX_ISUM); else Isum = max(Isum, -MAX_ISUM); u = kp*error + kd*Derror + ki*Isum;

@@stevenhong7099 Thank you very much.

Lol im not even in high school

alunos de ensino médio no Brasil nem sonham em fazer isso

Maybe you spend less money on buying a glue, and glue the ball to the tilt-ball system, lol. But my goal for the video just to show the control algorithm/technique, so not actually using a glue :)

@@stevenhong7099 it was a joke :)

@@yaminsiddiqui4690 Engineers have no humor, lol I just saw so many RUclips comments saying Boston Dynamics are fake and CGI, but I play with Spot every day. Sometimes hard to tell if people are joking or really don't understand the idea.

Simple PID, not using any advanced or modern control theory. I'm very surprised so many people watch this simplest video, as high school students can also do it.

@@stevenhong7099 I know PID but i like to see how people implement it. Linking D and I or ignoring D and just playing with P and I. I did slow HVAC control code.

@@ConsultingjoeOnline P is main driving force, I is correcting steady-state error, D is preventing overshoot. PI, ID, PD controllers are also commonly used in the industry depending on the applications. To be honest, not a lot of people using PID because too many parameters to tune. For this video, actually Integral controller itself performs the best. I did P, I, PI, PD, ID, PID controllers for this same tasks. Since it's running on a microcontroller, so I also have other tuning parameters such as sampling time, max_I_sum, max_d. Tuning process itself is very tedious. Although PID is the most used control technique but I don't like to work on those projects personally. I like to use LQR, MPC, robust control instead of PID.

@@stevenhong7099 Very cool. Yeah We would try keeping it simple and just have limits and an error val as I would just do temp or humidity control. Drive a demand percentage into a 3rd party system even.

@@ConsultingjoeOnline /*********************************************/ // implement the CONTROLLER (Gc) functions // PID controller /*********************************************/ Derror = error - last_error; last_error = error; Isum = Isum + error; if (Isum > 0.0) Isum = min(Isum, MAX_ISUM); else Isum = max(Isum, -MAX_ISUM); u = kp*error + kd*Derror + ki*Isum;

I used STL file to generate vertices in matlab.

Lol, although not much real content in this video. This is just a raw demo video for my professor, and even didn't add any caption or explanation to it. I put more effort into other videos though.

@@stevenhong7099 yes but is great, almost all videos related want the ball in the center.... But is kind of easy, there is a column at center!! So is a good stability point... But your approach go for different stability points and is a good control strategy and very interesting

Lol

Hello, depending on what you want to do. For this video, I'm just doing the forward kinematics, and as you can see in the video that I didn't implement the inverse kinematics. If you are doing the spherical wrist, you don't need Simulink for inverse kinematics calculations. However, since you mentioned that you were using offset wrist, it would require Simulink for nonlinear calculations (depending on the algorithm). In this video, I didn't use Simulink, but I do have Simulinks for another robot arm called ST Robotics R12 Robot, which has an offset wrist. However, the algorithm is based on the actual configuration of the robots. I don't think my codes would work for other configurations.

@@stevenhong7099 Nice, How do you put your cad into the matlab figure?

@@Alteregofr Using STL files to create vertices in MATLAB. However, it is a huge pain to rotate and translate the solid models to desired position.

@@stevenhong7099 so you use CAD file to demo this or what? I'm struggling because I use Peter Corke Toolbox and I can't make it understand the spherical wrist. I set my Stanford arm very similar to the puma560 preset but it understands puma560's wrist not mine. (I check it by robot.isspherical() ) Also, can you share the code? PLease :)

Yeah! Take Dr. Fisher's class. It's the best robotics course offered in Rose-Hulman!