Menlo Community Talk #4

追加: 2026-05-12

[00:00:31][clears throat] Hi everyone, welcome to community Friday by my law. I believe this is our fourth call. So um good morning or good evening depending on where you're dialing in from. So today we would like to start things a bit differently uh a bit of a structure. So I would like to um maybe structure this a bit uh mostly for our internal team actually.

[00:02:32]So today we would like to show you something that we shipped. So how did we make the robot better today? Uh stuck what are we still figuring out and uh we would also love to show you some some of the stuff that we've been working on because it's been very exciting two weeks at Menllo here. So we have a lot to show and for the agenda uh we will roughly go through this order. So first we'll have software and then we'll go through the hardware part and then some of the housekeeping some of the internal updates and career opportunities from in first uh I think I can pass over to Yuki. Right.

[00:03:15]>> Right. Yes. Um very excited to show you guys what we've been working on. Uh so we have been working on uh the manual platform. So so for the past few weeks and months you guys probably see a lot of updated on the physical ASM. So today we want to talk about the virtual one. Yeah. So we are currently developing the manual platform which supposed to help serve uh the virtual aimoff and then like everyone can basically own one as if it's like the real one and you can uh you know experiment with it or connect agent to it. So I just want to go through a bit on what we have done. Yeah. So uh upon logging in you can let's say select an environment. Today we'll probably go with the apartments and then let's open up the cockpit of Azimov.

[00:04:08]Yeah. So first when you guys go in see you can see the Asian lock on the right side and reasoning trace also. Then also all the email lock on the left side. So this is like all the data that we take from all the joint motor of the of the robot in the simulation and we just start streaming on the left side over here. Yeah. Just going to start the robot now.

[00:04:29]Yeah. So we actually done a lot of iteration on how to make this possible.

[00:04:34]Uh you probably see a lot of like demos of simulation in the browsers online already. So it's not like magical or anything but what we would really like to offer is like a much more high details much more complex scene uh in the browsers experience. Right? So to do that it's not very feasible to put that in the browsers. So a lot of this what you're seeing right now is actually the physic engine running uh in our servers only only the animation is done on the browser side right. So let let's go through a bit on the manual mode first.

[00:05:12]So actually you actually can control asim off manual mode as if it's like a game right. Let's try to give it a command. Just now I turn uh right now turn left. Let's walk through space a bit. Yeah, you can see on the right side on top there's like a firstp person view. This mimic where the asimoth camera going to be uh which is in the head. Yeah. And let's go back to starting point. Okay.

[00:05:42]Oops. Yeah. So what you guys seeing right now is like real local machine policy running on the Azimoff, right?

[00:05:46]Being trained by Azimov team which Yeah.

[00:05:49]will just provide like the full physicsity of it. And now the interesting part is we're going to try the Asian mode, right? Um it's it's still process like a work in progress, but it is quite interesting. Yeah. To see how an Asian would interact uh with the environment when they have embodiment, right? Uh hi. Can you see the doors on the left? Can you move to the left side of the doors over there and I try to close it?

[00:06:40]Yeah. So we have like an Asian loop running and then you're just trying to see from the first person view what what does it see what should we do next you can see all the reasoning state down here and then like the conversation lock is like on the right side right yeah but but but we're still working on the agent that that that part is hard like there's a lot of trial and errors is being done here it's like how many frame you want to send from the robot to the agent so that you know the history you know the context like or like how's the heartbeat going to be like all this designing like all this fine-tuning is what make kind of like the magical moment when an agent can control a robot. We're still working on it and in the future in the near future hopefully we would really really want to allow everyone in community to plug in their own Asian also so that they can cont so that you guys can experiment with like how an Asian would work in the environments uh like like in a physical environments right and like help us advance the state of uh Asian embodiment together. Yeah. So that's the state of the manual platform. Right now a lot of our R&D actually is spent on making sure that like we can serve a lot of people on our infrastructure like all the physic engine on our infrastructure while like doing the while like you can still see it on your browser and rendering nicely and smoothly and all that. Yeah. Okay. So that's all from me.

[00:08:10]Uh would love to have question from you guys.

[00:08:13]Uh sorry uh Techmax what you mean by frameworks? So everything you see here is actually built inhouse. Um the physic engine is mujokco but other than that everything else we built in house. So including the rendering part which we would love to be able to open source one day so like everyone can you know test out with it also. Yeah.

[00:08:37]Is there an API we can plug this into our own type of option? Yeah definitely.

[00:08:42]Uh so if you go to the manual mode over here and you go click down right you actually see we right now have keyboard we actually have a CLI so you can actually do like manual robot forward manual robot backward we we figure out like if we provide this it will be quite easy for you to plug in like a clock code or codeex or whatever to like sort of like just control robot and we also have like a menlo robot snapshot which take a snapshot of the screen sorry not a screen snapshot of the robot camera also so that like know where it is.

[00:09:13]Sorry, you can pan it around. So, yeah, I hope that's how does spatial navigation work. Uh, good question. Uh, we are still figuring that part out.

[00:09:21]Right now, we are sort of like genering generating like a mini map of the environment from the Muoko from from like from the like the Mujoko simulation and like use that as like kind of like uh a way to navigate around the map. But honestly, a lot of this is like up to research. That's why we would love to have more people joining our platform to test out their own spatial framework for like AI if that makes sense. Yeah.

[00:09:48]Uh no actually this is just pure feeding. Oh, so we are okay. Are you using slam for navigation? No. Uh we actually just purely feeding in the the the first person view camera you guys seeing there. The agent is figuring out the rest. Yeah, there's no slam at the moment, but we would love to be able to include that maybe like a basic depth anything or something feature that will just do segmentation and depth in in in the platform itself. Yeah.

[00:10:17]Yeah. You can see the the agents kind of fail here, but uh I just uh Yeah. Okay.

[00:10:24]Is there any other question?

[00:10:27]Yeah. Yeah. Yeah. That's the idea, right? uh we would love to have you guys plug your own Asian into the platforms, right? So, Orion is just a way for us to test our Asian control and it's like a way for everyone to uh like have a reference design, right? So, to how to like build Asian on our platform. Yeah.

[00:10:50]Okay. Yeah. Thanks for all the enthusiasm. I love to bring this to you guys soon. Yeah, we will love to bring that soon. Yeah.

[00:10:59]Please continue dropping your questions into the chat and we can um respond to you there. But I think next up we can move over to hardware team.

[00:11:13]>> All right. Thanks M.

[00:11:15]>> So the last time we left off I mean our last call we mentioned that we're going to release one. So now it's actually public. Um so far we've released the electrical schematics, the mechanical um files, the assets as well as the wiring diagrams for everything. Um as you can see here, we have small readmes in each of the sections to show you how to navigate the places as well. Um currently the manual does not have information on electrical data, but that's coming soon. TBA probably sometime in the middle of next week. Um, for the mechanical side, it's same as the legs over here, but we actually got a very helpful PR from the community to help consolidate some of the mechanical information. So, I'm not sure if Abdel Sluck is here in the call today, but thank you for this PR. I believe I understand what you're trying to do. I have some suggestions how we can improve it and I really want to merge this into the main branch.

[00:12:06]As for the road map and what we're having in the future, um, the Asimov API which is be going to be a primary way for you to control the robot. So based on Yuki's previous presentation, what you saw there in simulation, the exact same API can be used to control the real. So this will allow you to go from um developing applications in sim without having the robot and then directly deploying it. Of course, there might be some differences due to having real hardware, but the software API itself will be exactly the same. um for the local policy HD will share B soon and we're also working on a mobile app to make it easy for you to control the robot not just for tele operation but just for basic turning on monitoring logging um agent based um debugging of the robot itself is it on to you?

[00:12:51]Yes. Um I can share my screen.

[00:12:59]Yes. So I'll start off with the video of the current state of the locomotion policy. And this week we were just trying to really stress test it and test it on different surfaces. Test it without the crane to see what the failure modes will arise and what the edge cases will be with the locom motion policy. So as you can see we are trying to move it outside our office here.

[00:13:18]We're just trying to see does it get caught in the floor? How much is it leaning forward? How much drift is there? And so this is the locom motion policy. we are able to get it working on the robot after we got some hardware fixes. And as you can see, it's not perfect here. And the robot sometimes gets stuck or the robot's leaning forward. And so the main two issues we identified that we want to solve is one, the robot generally tends to lean forward. It could be that the policy learned a dynamic and we need to model something correctly to improve this. As well as you've not seen this video, but sidest stepping was an issue and so we were not getting determinism with sidest stepping. Sometimes it would just wildly swing its leg. And so one way we actually rectified this is we went back to the XML and looked at okay are we modeling the collision boxes properly since Moduk is all about collision boxes. It's just a contact physics solver. And so we realized that during a testing we had actually changed the initial pose of the robot but forgot to update the collision boxes. And you can see the arms are quite off the toes are actually quite off as well. the it gives the policy headroom to lean and like use its ankle ro so much but in reality it doesn't have that and so after we rectify the XML and went back to a more um aligned XML to the mesh um we were able to fix the sidest stepping issue because the model is able to actually know how much foot room it has to be able to bend its ankle so on so this solved the side stepping issue but we are still solving the um forward walking issue I think I go next.

[00:15:03]Or any questions about the previous part?

[00:15:11]Okay, then I just start first.

[00:15:25]sec. Okay. So, we have some update about the menu since um I noticed some people asking about the battery and the hand.

[00:15:34]So just to be clarified, we don't include hand and battery in our um asov one kit, but we do include some operation guide um and also like what do you need to connect to the battery and the hand for example. Um these are all the specs that you need to follow to build a battery that is suitable for the whole system including the mechanical dimensions and power capacities and also the charge current.

[00:16:04]Yeah, we do include a battery.

[00:16:09]Okay. Um uh we we'll update this soon. I think there are some changes about um shipping the battery and also we um include as this is also in the electrical files in the open source GitHub repository about how to operate and connect the motion control board and you can see the mapping of connecting through the picture here. So you can connect it yourself to um both the legs and like powers and etc. Uh and about hand we're not included here. So um we we we have like three options for you to connect um your hand. So like for for power you can connect either directly from the main battery but or you can connect through the waste motor port but just to be um just to um to to be considered that you have to have a converter to set down the power to the normal operating range. for example like 5 volts because um the voltage of the battery is like 48. So and also the communicate communication option we have three. So you can either through the motion control board which we will uh which you have to upload to own script to control the hand and also you can directly control through the Raspberry Pi or you can use whatever communication board you have. Um and either way you have to write your own script. Um, we do have some recommended hands from uh, units by hand and dex hand and also robotic and okra hand which they open source so you can take a look as a reference.

[00:17:49]Yeah, that's all for this update. Please um, share any feedback you have or you can just go through the assembly steps and let me know if there's any step that you're confused so I will add more information.

[00:18:04]>> Oh, Ariel, could you get back to the battery section? I think we had a typo there on on the manual saying the battery is not included. We're actually working to get certification right now so we can ship a battery along with the rest of the parts for the a charger will also be included. So out of the box what you will get is a battery and all the components required to assemble the so once you're done going through the assembly manual you have a working.

[00:18:25]Oh yeah, this um is included in the steps that like here you you can just install the same way as what the instructions is.

[00:18:36]>> Yeah. Does anyone have any questions about anything hardware related, any software? How do you update the URDF? So currently the URDF update is completely manual. Um it actually just two days ago we all sat around one computer and we moved each of the joints by like 0.01 01 until everything aligned perfectly.

[00:18:56]There's no good way to do this yet, but I hope as the robotics ecosystem matures, we get more automation tools for this. It doesn't make sense that I need to do this by hand.

[00:19:06]Um, actually, speaking of uh alignment, we're actually working on a way for you to home the robots. If any of you have used a 3D printer before, you know one of the first steps when a 3D boots up is that it moves all the axes to the end positions. And this allows the machine to figure out, oh, what is my the limits of my um my limits of freedom for the machine itself. And we want to have something like that for a humanoid. But getting it to work with a humanoid where you have some 360 joints, some joints like 270 degrees can be quite dangerous.

[00:19:37]So we're figuring out a way to do this with end stops and just software.

[00:19:45]Um so sorry just joining I've moved away from Ross 2 and using a custom ROS app.

[00:19:50]Are you guys on Rossu? No we are not on Ross too. Alive has a slightly different stance. We don't like using Ross for things because we feel that it doesn't have a yeah it's a trigger point. I'm being very politically correct here. But our stance on Ross is that we feel that it's an unnecessary abstraction if you actually understand systems programming and we do have some very good systems programmers on the team. So prefer to use the abstractions below Ross itself rather than using ROS direct. We don't really have anything personal against Ross but it's more of like only use what you need for the BMS. We are getting it along with our battery pack. So, it's um not really a custom BMS, but it's a BMS from a contract manufacturer we're working with. So, we're not actually building the battery pack ourselves. We don't really have the equipment or the knowledge to properly test a battery and a BMS. So, we are outsourcing this one, but we are working with the contract manufacturer to get it certified with the UN 38.3.

[00:20:57]Great job on the manual. The manual needs to have pictures for each of the parts because it is hard to see the IKEA diagrams without the part numbers.

[00:21:04]>> I see picture. [snorts] >> So like for each module you want to Oh yeah. Let me like for each module you you mean like you want to like export the view of all the parts you need for the structure instead of just the names.

[00:21:20]Is that what you meant?

[00:21:26]Yeah. We we can include that for sure.

[00:21:43]>> Yep. We will include that.

[00:21:58]Oh, yes. I hope your crowd sourcing goes very well.

[00:22:14]>> Okay. I don't think there's anything more from the hardware side. Mimi.

[00:22:19]>> Yes. Um guys, please continue dropping any questions in the chat or you can tag us on socials and then we'll get back to you. Um in the meantime, I think Herman has a lot of career opportunities to promote. Is that right, Herman?

[00:22:36]>> Yes, that's right. Uh yes. So, right now, let me share my screen.

[00:22:45]>> Yep. So, we have quite a lot of opportunities open. We're hiring a dev ops. We're hiring supply chain associates. We're hiring firmware engineers, a forward deployed research engineer, embedded systems engineer, platform engineer. We're also hiring marketing specialist. If you're based in Singapore or you're open to working remotely somewhere in USA, we're looking for you as well. Uh right at the top of our list is also a marketing intern, someone to help us storytell what we're doing here at Menlo Research. And there are quite a number of other opportunities that you can check out on the website. So I look forward to receiving all of the applications and having a chat with each of you. Thank you [cough] about what CAD we're using. So we we actually use Solid Works for all of our CAD. We are we have members in the team experience in Solid Works. That's why we chose it. Um, it's also got very good URDF converters and URDF HTML converter sorry not URF HTML converters but step assembly HTML converters which is very nice for visualization um about triging estop with software actually when the robot ships the only way to trigger it will be via software that's going to be the main way to control it so it'll be connected directly to the motion control board and there'll be a relay that you can trigger via GPI so the goal is that it'll be similar to like a car where like if you use most modern EVs, they usually have a companion app. And we're thinking about the robot in similar fashion.

[00:24:20]Um, and I I think legally we can't call it an estop because usually ESTO means it cuts all power. What we do have is that we have a relay switch off power to all the bothers, but the compute stays running. So the robot won't be able to operate, but you'll still be able to collect logs and find out what actually went wrong during the entire operation.

[00:24:43]the RL policy. What would you like to know about the RL policy?

[00:24:52]The RL policy we're currently using is just um a normal PO1. We will be releasing more information as to how we did CIS ID um how we solve the SIM drill in the coming weeks. But for the RL configuration or the model we are using, it's just the BPO for now.

[00:25:20]There was a question on whether the estop can be triggered via software. Um Tashan, I thought that was a really good one.

[00:25:28]>> Oh yeah, I I covered it. So to answer like very bluntly, yes, it will be able to be triggered via software. That's that's the whole goal.

[00:25:37]um go over what we use power distribution. Um so for power distribution because all the motors run on 48 volts, there's not much of an issue there. We just have a small bus out of XC connectors that allows you to distribute the power to all the limbs.

[00:25:50]And then for conversion, I'm currently working on a B converter to convert the power down. Um in the current design, it's offtheshelf version, but we would like to have our own so that's easier to integrate into the entire world. It's been giving us some trouble, so hopefully by the next community Friday, I'm able to share more.

[00:26:14]All right. Any more questions?

[00:26:24]>> Have we seen anyone do human or tele with a DCI yet?

[00:26:33]>> That's a very interesting question.

[00:26:36]um we have not tried it ourselves but I have seen some people online attempting it I think that's one of the NDM goals of Optimus and Neurolink so for for those of you who are unfamiliar with the acronym BCI stands for brain computer interface so the idea is that you'll you'll mount like electrode stuff onto your head or have an implant in your brain itself that allows you to control the robot it's it's this is talked about like the end game for tele operation you basically go from thought to action instantaneously Um so a question from Jay why do you guys communicate over a single board computer to Moz directly not communicating with microcontroller to increase bandwidth. Um so it's not really to increase bandwidth but currently each of our actuators and the reason why we call them actuators and not motors is that it already has a motor driver embedded into it. So there is multiple microcontrollers around the robot. I mean 25 to be exact for each of the joints. So we don't really need that high frequency control loop that you get from a microcontroller. Instead, what we opted for is to use a full single computer that can run Linux. So give us some development velocity, makes it easier to OTAA and just development ease. And Linux drivers are pretty nice, especially when it comes to spy can and socket can.

[00:28:01]>> Yeah.

[00:28:02]BCI teleop would indeed be awesome.

[00:28:13]>> Yeah. Um, we've actually been trying to get I mean not trying we have gotten VR tele operation to work using the robot but one of the problems we faced is that I mean it's the same problem with VR games right is that you can get good hand operation if you're standing in place but how do you handle movement like squatting down picking up stuff and without you actually displacing in real life. So you need quite a big rig to do the whole full tele operation and it's quite actually quite uncomfortable to operate one because you're always slanting and stuff. I mean, you can always use a joystick, but your joystick is not getting full freedom, right? It's kind of like playing on a controller.

[00:29:01]Let me see if there are any more comments. Um, is Mando here with us today?

[00:29:08]Mando and Grafto, they've been doing a lot of work to actually fabricate the robot. We we were very impressed by your workshop. Mando >> Mando is in the chat. He's watching us.

[00:29:18]But you're a legend inside the company.

[00:29:20]Like we talk about you. We're so excited. [laughter] >> Oh, okay. Hello there, Bando.

[00:29:44]So Babalaf is saying that Colonel will be interesting in partnering with you guys. He was very supportive in my three hacks project. Oh, that's amazing.

[00:29:52]We'll definitely reach out or if they could reach out to us either way.

[00:29:58]Uh Richard Franklin, I think you can get a squad with head tracking on the headset, right? Yeah, but it's also like the dynamic movement like you're going down and like reaching for something far away. It's I think it's definitely possible to do. It's just it's a software engineering problem which I like because it means that it's possible to do today. It's just a matter of how creative you want to be about it.

[00:30:24]So ballerdude is sharing his setup. Uh I'm using 8 bit do mobile controller connected to my phone that sends teleop commands via websocket connection from app to the Jetson board. Just an idea for you guys. Yeah, I think this is a pretty elegant way to go about it. Um currently what we're doing is that we just use keyboard operation, but of course you can connect any kind of hardware interface you want. And since we use live kit as our main way of bridging the robot to the outside world, is this as simple as writing a life kit publisher.

[00:31:07]Okay, I don't think we have any more comments or questions coming in. So again, please continue pinging us on socials, on our discords, and we continue our conversations there. But um otherwise, thank you for joining today.

[00:31:21]Uh we love to have you here to watch us and follow us. So thank you for your time.