"Robotics Today - A series of technical talks" is a virtual robotics seminar series. The goal of the series is to bring the robotics community together during these challenging times. The seminars are scheduled on Fridays at 3PM EDT (12AM PDT) and are open to the public. The format of the seminar consists of a technical talk live captioned and streamed via Web (roboticstoday.github.io/watch.html) and Twitter (@RoboticsSeminar), followed by an interactive discussion between the speaker and a panel of faculty, postdocs, and students that will moderate audience questions.
But who are you selling the product to? What will be its purpose? Wonderful, but how many people need a robot with these skills. Learn from Elon, the robot needs AI. It needs to interact with people to be helpful. Enough with the jumping.
Wow, this was so inspiring and interesting to watch. The engineering behind this product is just beautiful. Watching the people behind the product, i feel this truly is only the beginning.
CBF: Control Barrier Function CLF: Control Lyapunov Function QP: Quadratic Programing RL: Reinforcement Learning GP: Gaussian Process regression SOCP: Second Order Cone Program HZD: Hybrid Zero Dynamic
I wanna buy one of these so i can get it to goto the store and buy my beer, cigarettes, and lottery tickets :) Ok everyone don"t get any ideas that was suppose to be a point of humor not an actual application :) unless you want your bot to get stolen and sold for drugs :) I'll just wait and order from the robot chop shop on tor :)
I have a MS CS from the big state U. And I have not the faintest clue how this thing can separate reflected image in the bridge's shiny inside walls from real images of obstacles. You can't just let AI do it. AI needs geometric clues too. The reflected images are all over the place in the view.
When blindly using deep learning networks to solve problems, without a fundamental understanding of the problem being solved, mostly leads to "garbage in, garbage out". It's refreshing to see companies embrace deep learning, but using a modular, first principles approach, as opposed to just blindly hitting the problem with a large deep learning hammer.
13:56 was certainly a white-knuckle moment but definitely impressive. Nice job handling the thin power lines with high contrast background. Especially 14:39
I think the focus is way too specific. Until Neuralink becomes available to us humans, the machines will never really be able to read our (a users) mind and that should be okay! I think generally we expect machines around us to have more discrete action spaces than continuous ones? I don't know how I feel about being surrounded by smart machines getting all creative and learning new things all about and around me. I'd say people (users) need to be able to modify their own policies and be able to adjust to the range of capabilities the machines have that they use regularly and that would help the relationship just as much as a constantly changing robot that would perhaps become considered unstable and unpredictable. If anything they just need to be able to start off more general in new environments and be able to adapt and acclimate their policies based on the environment they've been put in and really shouldn't have a bunch of expectations and learning it would have to undo. So basically more general reward functions that can learn from it's environment would be more acceptable than being all pretrained to somewhere it wasn't exactly deployed to... PS I really liked this presentation! Way better than the last 5 or so I played...
Bad approach, there is no learning. Engineers can't code everything because they don't know all the dynamic and perceptual things they do, most human behavior has not even have a name. Also MPC won't cut it in the search period of that learning. Beautiful toy, no vision, I mean the team.
This makes me angry. How arrogant is telling that their approach is bad. They have decades of experience building walking machines. They have shown amazing results that no other people has even matched. And then, you, a nobody, tells them they are wrong. Shut up and listen so you can learn something
@@Robotmorales They didn't show amazing things, your threshold is very low, MPC is 41 years old now. That puppet is not customizable for anything other than its own advertisement. And the fact that they are bad at their jobs & their vision is not correlated with me and my anonymous account. I can explain why they are wrong, but I doubt you have the capacity to acquire that knowledge. I am open if you are open.
@@444haluk The fact you only focus on MPC demonstrates that you have missed 99% of their work. You might think it’s a toy, but it’s a real hardware toy that works effectively in the real physical world. It’s yet to see wether a ML based approach can provide a small part of the same performance. I still believe that it is pretty arrogant to qualify their job as bad or their vision as wrong with such shallow arguments as yours. And don’t hesitate to give more detailed arguments. I assure you that i’ll be able to understand them.
@@Robotmorales 1) Embodied AI: before you aim to work on some area, you need to define things and they need to be water-proof. You cannot just do a body without the ability to learn because it will not survive. MPC is not a learning algorithm, it's a static model predictive algorithm. It cannot model anything that you didn't put your time into. Will you write everything when Atlas gets a hand? And when it tries to lift? When it tries to hold onto something? Because with MPC, you have to. There is no "let it figure out". 2)Motivation: An agent has to have a motivation, else there is no difference between doing one type of action and another. Humans call this phenomena "indifference". 3)Hard coded high level tasks: Now you have to assemble those micro tasks that you hardcoded, into a macro tasks that you hard coded and you spend millions into one specific map that you have to record beforehand. Change the map and Atlas cannot plan fast enough. 4)Hard robotics: there is no need to make a body if you are going to use hard cold actuators, because not only it will not dampen the external forces for its own sake, it can also hurt everyone around it and everything it carries. 5)Who learns: Engineers learn, not Atlas. Engineers are after quick money, when the combinations catches up to them, they will only burn money. 6)Impossible=possible?: MPC cannot determine the difference between possible and impossible, that's why it is not suitable for real world planning, if you consider every object in every possible x,y,z,phi,theta,psi, you gonna have a hard time. 7)active vision: it uses lidar like sensing, which cannot be used under sun efficiently. It is basically blind outdoor. Bonus: it is 41 year old algorithm, not exciting. they are basically 30-35 years late to implement it and move on.
@@444haluk I see new what are the requirements of your vision: 1) Embodied intelligence. I agree. This is a must if we expect robots able to adapt to untrained cases and engineering
Great as always Davide. There is no doubt that event cameras can be revolutionary, but the idea of replacing the traditional VIO+MPC with an End-to-End approach was mind-blowing. I'm still thinking of the decades of efforts behind complex visual odometry and SLAM algorithms beaten by DNNs. The old belief of drift is somehow disappeared. I want to add that there will be no issues regarding the efficiency of DNNs in the future. Thousands of companies are investing in TinyML. More and more companies are producing TPUs and GPUs. The Intel Realsense you have used is an example of this. How much change we are going to see in the future? Who can find something more appealing and exciting than science? I'm getting almost mad!!!
Interesting video. Thanks for sharing so much information in details about Curiosity. It will be interesting to see how the helicopter 'Ingenuity' continues it's journey using visual navigation.
Great talk and impressive work. The citations at ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-EGABAx52GKI.html, are slightly occluded. Can we have the full list to try and understand better the math behind what Scott was explaining?
This is gold, I want to thank Professor Sidd for bringing so many state of art concepts in Motion Planning in one presentation. I will definitely try to read a few of those references. Thanks, Robotics Today for bringing this to us.
Scott Kuindersma: If it would be appropriate, I would urge you to get deep into Machine Learning. I thought the movements of Atlas were already controlled by an AI, because it looks like a very suitable task. Machine learning is so powerful it is mindbending. It looks very much like using machine learning already. Spontaneously, I would think of optimizing movements to use minimal energy by learning on the model (I expect that would create natural movements), walking on ground that can not be modeled, like irregular unstable stones with sizes similar to the feet. You would just all measured joint angles, bending forces, motor currents and hydraulic pressures and acceleration of every part of the whole body into a network, and let it learn to walk. (It can handle falling a couple of thousand times, I think). Of course, it can happen that machine learning also solves what you did before, with detailed modeling. It's possible - it happened to me. This is an impressive example of robot controll by machine learning: A robot hand with similar joints like a human hand learns to solve the Rubiks cube. Learning how to rotate the pieces is the simple part. Learning how to manipulate a Rubiks cube in one hand is not simple. openai.com/blog/solving-rubiks-cube/
According to Marc Raibert (chairman of Boston Dynamics) in the talk he gave with the Alan Turing Institute recently (video on youtube), the Atlas team is pushing hard toward using ML for controlling their robot.
Prof. John Leonard has really pointed out an issue of SLAM. There're always rumors in academics claiming SLAM is a solved problem while the robustness of algorithm seems always been confined in Lab or demo scenarios. The videos shown by Prof. Andrew Davison are all with steady camera movements. What if they encounter bumpings or sudden moves or perhaps bad illuminations? I'm not criticizing his works. They are great and inspiring. Just that it'll be cool if the robustness of SLAM can be studied. Think of what we can work on in such a direction.
