BTW, for those who don't follow Tesla, there're rumors before AI Day that Elon may unveil something like this, and Tesla may be working with with UCLA's Professor Dennis Hong, who specializes in Humanoids & Bipedal Robots and runs the Robotics & Mechanisms Laboratory (RoMeLa), you can see some of the lab's creations here. See this article (published before AI Day) for the possible Hong connection.So don't assume Tesla is working on this starting from zero, they may be starting from the best we have and improving it.
Same Topic. Different aspect!In the presentation, Tesla's ability to simulate situations was a key strength. This can be applied to Mars or eh Moon.In the context of robots using implementations of the same system it should somewhat follow their current methods for training FSD cars, however it will be applied to dusting solar panels, plugging in cables connecting pipes, bulldozing regolith, etc. When stuck, the returned video etc can be used to improve the simulation, retrain the net, and send an upload back to mars, which will help improve that task for all robots of the same model. It wont, just be reposition 5cm to the right and retry the same procedure!
A space / moon rated robot would need to be capable of operating in a vacuum and a very dusty environment. Essentially really robust with the same lubricant less joints like the Canada arm and parts that have none or have completely sealed volatiles. One of the barriers to making such robots is the cost of the one off unique components. The potential to bring such costs down is mass manufacture, any robot that is rated for space would be and exceedingly robust robot if used in the relatively benign environment on Earth.
Humanoid robots aren't going to be anywhere close to useful on Mars for a long time.I would envision picking a location that is easily navigated by non-humanoid machines, and use them remotely to carry out as much work as you can. Time would be better spent on non-humanoid machine that have specialised appendages for common tasks, or which have more generalised appendages that could handle a variety of tools, but have them controlled by humans (in the same vein as the tele-surgery that can be performed now).
Humanoid robots are the right choice only in environments built for humans where it is hard to fit a specialised chassis. Existing Earth buildings are built for humans. Mars is not. There is no good reason to use humanoid robots on Mars when you could use robots more suited to that environment.
I’m not sure we are getting the full implications of DOJO/FSD as it applies to a humanoid robot.If I were to drill a hole I would first select the drill. In just a few moments I would,a) Select Size.b) Select Geometry of Cutting Edges. (Flat Bottom, 118deg, etc.)c) Are cutting edges sharp?d) Are cutting edges chipped?e) Are cutting edges true? (One edge longer than the other or at a different angle, IOW badly sharpened.)f) Are the flutes tapered due to wear?g) Is the drill bent?h) And so on.If it passed my inspection I would use it, if not I would sharpen it or discard it.Hopefully I would get my hand drill and produce a good hole.Now lets say I wish to use a CNC Milling Machine (Robot) to drill the hole.I still need to do all of the above! But now I load it into the CNC with a few parameters, I have to set the following,Must Have Parameters1) Length overallOptional Parameters2) Diameter3) Angle of cutting edges (118deg)4) Length of flutes5) Maximum pressure allowable to push the drill through the material6) Maximum cutting time of tool7) And so onIf we program the machine with these optional parameters in mind we can build in some “Artificial Intelligence”. For example perhaps in the past we have noticed this drill doing this task breaking at the 100 hours of cutting time mark. We can set the parameter to say 80 hours and hopefully never break a drill again. But does the CNC know the drill is blunt like we do? No. It is “Artificial Intelligence”. It’s not real. And of course this doesn’t take into account extremely unlikely scenarios that you have not accounted for in your parameters and program. Notice here that there are only a very few parameters.Now lets say we instruct our Humanoid Robot to drill a hole. Using its FSD system, that has been “Trained” by the DOJO system, it selects a drill just like I would do! It SEES the drill. It examines the drill closely. It decides to use it as is, sharpen it first or discard it. DOJO has trained it on how to set up the drill. DOJO has trained it on how to use the drill. We need DOJO because this training results in hundreds/thousands/millions of parameters and scenarios. One of these parameters would be the pressure required to push the drill through the material. Our Humanoid Robot after drilling many holes senses the increased pressure required. Stops and examines the drill. “Sees” that the drill is blunt. Sharpens the drill and continues. But does the Humanoid Robot know the drill is blunt like we do? No. It is “Artificial Intelligence”.DOJO comes into play by simulating/analyzing all of the different things that can happen when drilling a hole and developing a scenario for each.FSD comes into play by selecting the correct scenario based on what it is “seeing” in real time.OK. So hopefully our Humanoid Robot can now drill holes. All we need to do now is train it for all of the other tasks we want it to do! Who knows how many parameters and scenarios are required!However if the Humanoid (FSD) Robot comes across a scenario that DOJO hasn’t allowed for it will stop. Even with this limitation I believe that Humanoid Robots will be very helpful on the Moon and Mars and given DOJOs ability to learn over time by examining an ever growing library of real world examples it will only get better over time.
Not really true. Humans are that shape for a reason - it’s an efficient shape for doing general purpose tasks. Human shaped (a more generally, animal shaped) robots will be useful anywhere where you want them to be able to solve a variety of problems, or problems they’ve not encountered before.
