cbames89

Sweet, thanks for sharing your experience!

UR website[0] seems to suggest +- 10N (a kilo resting on your neck) with their stock implementation. I imagine you're going to want something better than that.

What Force/Torque sensor do you use? Some of them are temperature sensitive and require a good warm-up period before they can be used to achieve good performance.

[0]https://www.universal-robots.com/how-tos-and-faqs/how-to/ur-...

Great improvement since the last video!

Are you just using Jacobian transpose for force control on the neck? I ask because it seems a bit stiff, the red marks on the neck of the test subject suggested to me that maybe there's a lot of noise in your force sensing, or at least unmodeled behavior.

Due to the slow growth of test subject hair, have you found 6$ to be cheap enough to keep your test cycles short enough?

The kinematic configuration of the arm looks odd, did you pick this to maximize rank of the Jacobian or the manipulability matrix?

I completely agree, the alphago paper is more like the working notes of a team that just accomplished something. The alphazero paper is a more careful analysis about the theoretical underpinnings and paints a clearer picture.

In North Carolina, it's very easy to get proper vegetables.

It could provide the motive to begin a(an even more?) massive automation push within the U.S. That could have long-term negative consequences for China.

One of the labs that I visit frequently has Beams for people to use. They've been on a multi-year kick to make the lab "Beam-able." That means, no extension cords, no floor outlets, no ramps, or edges on the floor of any kind. While in theory the ability to move around should make the interaction better, it seems that the mobility is so poor, that people generally get to one location and then just stop trying because it took so much effort to get there. It's also hugely distracting if someone is in a beam trying to get a better view because the inevitably run into something, or get stuck, and then need help. Elevators also turn out to be pretty big traps.

This is pretty similar to my experience, but I also have two kids both born during graduate school. Their addition to our family will delay my graduation a bit, but they're worth it :)

Byron Boot's lab at Georgia Tech does a lot of interesting work with larger-scale R/C vehicles. That being said, by adding more degrees of freedom you're making the configuration space of the robot higher. In many cases, higher dimension configuration spaces are more difficult for traditional sampling-based approaches. Learning approaches may be able to bias their solutions to avoid large sections of this configuration space.

Sorry to be a bit pedantic, but reality exists independent of belief or acceptance. However, a belief in whether or not a person can change that reality at some future point, I think is a more accurate description of the quality you're describing.

Maybe at the highest end, they're 3000, but a quick search of apartments.com shows there are still plenty of nice apartments to be had in the 1000-1500 range.

I think a more nuanced way of defining iii is that the intelligence an agent is capable of is limited by the extent to which it is capable of perceiving and interacting with an environment.

At one level this has to be true, if it weren't I could plop a black box on the table and say I've invented AGI it just can't interact with anyone, and you would have no recourse but accept my statement. We must necessarily define intelligence as the interaction that the agent can perform with some environment otherwise we'd have no way to know of its intelligence.

I've actually built a Dexter with the guys from HD. One of the beauties of that arm is its ability to force sensing inherently, at least the version I have. I doubt you'd need additional force sensing capabilities on top of what you get from Dexter.

However, I suspect you're near the arm's payload limit with the clipper setup. Feel free to reach out. I get to the Bay Area every semester or so.

Interesting project! I've thought about this one before, mostly due to Jimmy Neutron.

What arm are you using? Also what depth sensor?

A few suggestions:

It would appear that you're taking a pretty common approach to robot manipulation, sense, plan, "close your eyes" and then act. This works for stationary objects when you have high accuracy control and high accuracy sensing. Unless you're envisioning strapping people's heads down, your target is going to move. This will necessitate a closed-loop approach.

A closed-loop approach here would benefit from force control at the end effector. ATI[1] is a good source for 6-axis load cells and they've recently come out with smaller form factor load cells.

A second useful sensor would be a Time-of-Flight depth sensor. The old PMD Camboard Nano used to get up to 90 FPS and a close effective range (if I recall correctly ~3-4"). The time of flight sensors can be a bit noisy, but I've written up some of the smoothing approaches I've taken [2].

Good luck!

[1] https://www.ati-ia.com/

[2] http://www.cbames.com/publications

I learned a similar lesson from working on robots. At first I attempted to devise methods for Robonaut 2 to do things the way I do because it was designed to be like me. It was missing little things that made my approaches unfeasible, and it was infuriating. At that point I decided it only made sense to make methods that allow the agent to discover its own behaviors, because its merkwelt and my own will never be the same.

This is a (perhaps short-term) failure of first principles reasoning. Elon's known to favor thinking about things from first principles, and there's no theoretical reason that vision can't work. However, technical limitations might take a while to overcome.

A clock, a piece of paper and a pen.

I think an important aspect in this post is the idea of a maximum amount of time for a particular section. I've had some success using this to combat perfectionist tendencies. In addition, im currently going through grad school with 2 kids and reading about others who have made it is always good for maintaining the "can do" attitude.

One of the founders wrote a FPGA language. I suspect it's the tool the team knows the best, but may not be the most price optimized. I think that's fine for now, especially since it's still one of the most affordable arms available.

Ya recalibration is definitely a necessary aspect of this approach if the optical reading changes. For example if IR light levels change.

Like optical flow in that features are used, not like optical flow in that there's one standard by which all measurements can be compared to.

The FPGA enables higher sampling rates, which allows for the finer resolution when paired with a lookup table which contains systematic error.

There's just one encoder per joint. The encoder disk is 3D printed and the cruft from the print is used to refine the encoder ticks. Instead of fighting noise they save a map of all the systematic noise and then can use a lookup table to localize.

*I went to Vegas and built an arm with the founders.

I believe they have a version called electron which is doing smd placement.

I have a Dexter. They have harmonic drives from China which were deemed out of spec for normal usecases but are fine for the 3kg realm. Their latest version is running a custom cycloidal gear system.

Sure, my statement was based on rumors that I heard about Rethink taking a look at Robonaut with the covers off and the actuators looking eerily similar. But yes, Series elastic actuators weren't invented at NASA.

As a data point. The original Valkyrie cost ~12.5 million. Thats hardware and software dev over about 3 years. I'm not positive about the subsequent copies, but I've heard they're in million to 2 million.

You can get a spot mini for ~30k according to Raibert.

I think limiting our analysis to one particular form of robot is probably not a good idea. As we're still trying to figure out which form is the best for whatever that "killer app" is in robotics.

Also one reason to not turn a kitchen into a robot is generality of solution. If you can have 1 robot work in 3 kitchens is cheaper than retrofitting the 3 kitchens.

All of that being said I generally agree with your sentiment, and I think we'll see robots move into places they're needed/bring more additional value than general house chores. For example, restaurants. Where they can be faster, safer with food, but still general enough to put cheese on one burger and not the next without an engineer coming in after installation.

A lot of Robonaut's actuator technology ended up in Baxter. Now Rethink's dead, but it did start the whole cobots sector. In many respects the cobots from UR are closer to the model T then 60s era robotics. 60s era robotics is more like the tracked steam shovel.

Grep Beat is a local newsgroup that's keeping track of startups.

I looked on Glassdoor not too long ago for AI/ML and there were probably 20-30 postings.