StealthP3D
Well-Known Member
If you going to make a career on Youtube, just make sure you have orders of magnitude more success than I.
View attachment 860048
Nice! It looks like you have targeted the "smooth jazz" genre of robotics engineering primers. I could especially relate when you talked about how having multiple joints provided you with more freedom. I know I always felt more limited when I only had one joint left, so I felt like I was really starting to get the big picture!
All kidding aside, I was disappointed while watching Tesla's presentation to see a non-athlete being used to model human motion (14:45 in your video). The guy with the kinematic sensors strapped to his body should be an athlete with an exceptionally beautiful understanding of how to move with efficiency and grace. Using a middle-aged desk-bound geek for this purpose seems like a major blunder (no offense to that particular employee but everyone has roles they are particularly suited to). The creation of data (for example the kinematic sensor data) that will be used by many people in the development process should be sacred and of the highest possible quality to get the most out of it.
I am also wondering about your thoughts on bi-ped balance. I think you mentioned that the robot must keep track of its center of mass because unexpected perturbations could upset this. I envisioned robots blowing over in a strong gust of wind, something that would make the robot look quaint and useless. I'm curious if you have an opinion on how Tesla needs to detect perturbations. Obviously, humans use multiple means including nerves in the foot to measure pressure changes, nerves in other areas to detect pressure (for example a shove to the shoulder or wind pressing on the windward side of the body), the inner ear, vision (for example of the horizon), and perhaps a sub-conscious awareness of the movement of our center of mass.
I imagine if a robot could track the movement of its center of mass accurately enough, and at a high enough frequency, combined with local accelerometers located at key areas of the robot, that would allow the outside forces that were acting upon the robot to be calculated accurately in order to respond appropriately (counter-act the outside forces). Of course, it would be possible to supplement balance awareness with many other sensors including vision and pressure sensors. What do you think is the best strategy for Optimus? Which inputs should be the primary inputs? Should secondary inputs be fused with the primary inputs or used only as a confirmation? It seems like the same conundrum Tesla faced with vision and LIDAR, what to do when the inputs disagreed with each other. But intuition tells me balance needs to be more immediate, operating at a higher frequency, for it to be robust, to rely on vision alone. I don't think vision-based balance (combined with self-awareness of body positioning) would have enough resolution to be used for balance corrections.
In short, I can see how the computations required to maintain superior balance in challenging environments and/or conditions could consume a lot of processing energy but, on the flip side, not having superior balance could severely reduce the utility of a robot by making it vulnerable to expensive damage in challenging conditions and neuter its utility. It comes down to risk/reward. The owner of the robot does not want to take (for example) a 0.1% chance that it could sustain $200 in damage in order to complete a short task that is only worth $0.20. Because even though that looks like a "push", it's inconvenient and takes your robot out of service until it can be repaired.
How do you think the robot best achieves robust and versatile balance without consuming excessive compute power?
The investor in me senses that the Optimus development program might be too lean of an operation for the optimal return on investment over time. They could let the "git 'er done quick" teams continue their work while having parallel teams taking a longer-term approach that is already developing the next generation designs (without actually building them). The learnings from the quick teams can be shared with the longer-term teams as they go, with the longer-term teams developing the next platform (or a higher-end, more capable, heavier duty platform). I think it makes sense to have an inexpensive lighter-weight, weaker platform, and a more robust, more capable and more expensive heavier platform. I thought the initial Optimus would be even flimsier/cheaper than what I saw but it looks like they are building a robust and durable high-quality robot. I'm sure with time the platform will split into two models (cheap and light, heavier and more robust), but it probably makes sense to develop them in parallel from the beginning (or very near the beginning). So, I hope the AI day recruiting effort was a raging success. When the economy falters, and the economic future looks most uncertain, that is when recruiting can be most successful. The potential reward of highly versatile humanoid robots is so lucrative I think it makes sense for Tesla to have "layered" teams working on this in parallel.
Last edited: