Skip to playerSkip to main content
  • 22 hours ago
Kangaroo is an AI-powered humanoid robot showcasing advanced reinforcement learning and imitation learning capabilities. Leveraging GPU-accelerated training in simulation and seamless sim-to-real transfer of learned policies, the robot achieves fast and stable walking, robust balance and push recovery, dynamic jumping, and expressive full-body motion generation. The system combines agile locomotion with imitation-based behaviors such as dancing, demonstrating recent advances in embodied AI, robot control, autonomous mobility, and real-time motion adaptation in complex environments. To showcase these advances, the Dynamo-Grade joint lab team will conducts high-impact live demonstrations, positioning Kangaroo as their core bipedal platform for next-generation humanoid robotics.
Transcript
00:00Okay, hello. Hi everyone and thank you Eric for the kind introduction. My name is Albert Tissot. I'm from Pile
00:08Robotics. I'm here with Olivier Stas, a researcher from LAS NRS, head of the robotics department, as well as with
00:15my colleague Maximilien.
00:17We'll be operating today the kangaroo robot, and we'll showcase besides the kangaroo that you see on stage, our humanoid
00:26robot, we'll show how the collaboration between labs, product labs and research with private companies helps make better products for
00:37research and industry. So yeah, I'll leave the floor to Olivier to proceed.
00:43So thank you. So to give a bit of context, we have been working with PAL and PAL France for
00:4910 years. Actually, PAL France was created to work with us. And as you can see, in 2016, we have
00:57been specifying one of the first humanoid robots targeted for industrial application.
01:03What does that mean compared to the smaller humanoid robots that you have seen that were aiming for, for example,
01:09going to factories, handling drillers, which are heavyweight, with the arm stretch.
01:16Okay. And this means that you need really powerful arms to do such kind of things and powerful legs in
01:22order to handle this.
01:23Now, as you may have seen, all these super nice small robots, which are super fancy. The problems, as if
01:30you follow a bit the trend of agility robotics, one of the first humanoid robots, which has been tested on
01:34logistics, there is a huge, huge technological gap to go to a humanoid-sized robot.
01:39And this is the reason that we have this crane over there to protect the robots. And for this, we
01:44have been working in our first European project called Memory of Motion, where we extended some of the techniques that
01:53were used by Boston Dynamics for the Atlas hydraulic robot, making the backflips.
01:58Actually, when they were doing scientific presentation, they were citing five papers, two of them were coming from our robot,
02:05our team. And we started to do a torque control robot, and this is one of the first commercial torque
02:12control available robot, which is make it natural to interact physically with a robot.
02:18We are focusing for us on the physical part. Then we got a store of Europe for this project. We
02:26started a new project, which is Agimus, targeted, for example, for mobile manipulator.
02:32There is one on the stand of PAL France that you can handle it on the same kind of technology,
02:37cheaper, and with which now we can do debris, something that, again, factories want to do, you know, removing this
02:44small metallic part after making drilling.
02:46And we're able to do this, handling planning, real-time control, torque control, you can interact with the system, and
02:54we're able to go inside some of the complex pieces of Airbus in order to do this kind of deboring.
02:59And at the same time, we created, thanks to Jan Air, a LabCom with PAL France to actually develop these
03:06new robots that I'm going to talk about later on, and we gave a lot of feedback, but it's taking
03:11a lot of time, a lot of iteration, in order to achieve this kind of results.
03:15And I'll let Albert continue.
03:18Okay, so we'll get a bit more in-depth in the products we have in PAL Robotics, and how the
03:26collaboration really helped us push the product forward.
03:29So what you see here on the slide is our Tiago Pro robot.
03:34Tiago is an acronym, it stands for Take It and Go, which is the type of use case it was
03:38developed for, which is a use case we find a lot in the industry,
03:43you know, to migrate one object, taking from point A to point B, it's a lot of manual work that
03:50is actually in industries that we aim to address with this robot.
03:55So some specificities of this robot, it has two arms with seven degrees of freedom each, with joints that are
04:05controllable in torque.
04:06As Olivier explained before, we've worked a lot on torque control with LAS, mainly with the Agimus project, with which
04:16we have actually implemented this type of control on this robot,
04:21which makes the control of this type of robot compliant, and then inherently more safe.
04:28If the arms are also equipped with automatic brakes that trigger in case of unwanted emergency stops, which then the
04:40behavior that the robot will showcase when it stops,
04:45the arms will freeze in place, as compared to other robots where the arms might fall, as this might be
04:54the case maybe with some humanoid robots.
04:56So in this case, the road freeze in place, without harming humans around, the environment, equipment, or even the robot
05:05itself.
05:07Quick word about the base, it's equipped with two lidars, which make it aware also of humans in its vicinity,
05:19with cameras, with cameras, which allow it to deploy flam navigation algorithms for robust navigation.
05:27So all these technologies, let's say, are different layers also of safety, which make this robot certifiable in safety in
05:37certain specific applications.
05:40Moving on to the software stack of this robot, as all PAL robots are integrated in ROS2.
05:52We are a company that's really committed to open source, to really help push the research forward.
06:04So, yeah, we maintain the repositories, and if you use one of these robots, you'll have all the functionalities right
06:13off the bat for you to develop your own applications
06:16or conduct your own research, if you use one of our robots.
06:23Quick word also on the human robot interaction capabilities of this robot.
06:29PAL really has also had a strong commitment to social robotics since the beginning.
06:38So these robots, as all PAL robots, besides being focused on applications, research, it's also focused on the capabilities to
06:50interact with humans.
06:51So we strongly believe that the way robotics has to be rolled out, the human has to be put in
06:58the center,
06:58so then the robots need interaction capabilities with the humans.
07:05We also have developed another cross-fragatory modules in power robotics that can be equipped on this robot to interact
07:16with humans.
07:18Just to finish on this robot, it can be used to develop research on any types of research,
07:27but also, yeah, on industry applications, as we were saying earlier.
07:32Okay.
07:33Moving on to the...
07:34We will now have a demonstration of the robot.
07:37So this robot has a particularity, as you can see.
07:40It has an actuation system very different than what you can see on the Chinese, actually, robots.
07:45So the actuation system that you can see on the Chinese robot comes from Professor Sung Bae Kim from MIT
07:52in 2016.
07:54And one of the main problems you have with this kind of motors is that it tends to heat up
07:57very quickly.
07:59Okay.
07:59So now the interesting part with this new kind of a system is that you have differential motors here,
08:04which are less heating up, having stronger charge, mostly because you have a very parallel design from the top.
08:13And most of the masses is on this robot.
08:16And here you have a heavy charge.
08:18I think we can go up to 60 kilograms.
08:20And in the frame of the T-Rex project funded by the France 2030 program, we're going to try to
08:27put more heavy loads and see how we can actually extend the test on this robot.
08:34So here, the walking is using the same kind of technology that you have seen, again, on the Chinese robot,
08:41using reinforcement learning.
08:43It worked beautifully in one year, where they battled for two or three years using classical techniques.
08:48And actually, one of the students implementing these kind of RL techniques just published a paper with one of the
08:56famous founding fathers of RL, Peter Abil, from Berkeley.
09:00So here you can deploy real technology.
09:02You can deploy your own system because it's open source.
09:06And we, as the research community, will continue.
09:10You see testing, again, this kind of powerful structure because one of the problems with these heavy robots is when
09:17they fall down,
09:18very often the impact is cubic to the size of the robot.
09:22So most of the times, when a full-size robot fails, it's broken.
09:26And here, we believe that with this kind of legs architecture, we are far much, much stronger.
09:33And for us, the point is that to develop, again, new technology partners with a partnership on different kind of
09:41companies.
09:42So we have an ongoing agreement with Wondercraft.
09:48We have, in our lab, an H1, so we are trying to compare what we can do with Power Robotics
09:54and other companies.
09:57But again, what we like with Power Robotics is they are committed to open source, which gives feedback to the
10:02community at the whole.
10:04And I think this concludes our presentation.
10:06Thank you very much for your attention.
10:10Thank you, everyone.
10:11So, yeah, we have really short time to cram in a lot of information, so I really invite you all
10:15to come to our booth if you want to talk more in-depth about, yeah, the robots, the applications.
10:21Yeah.
10:22Thank you, everyone.
10:23Thank you very much.
Comments

Recommended