Extreme parkour with legged robots • Hacker News Board

Re: Extreme parkour with legged robots

#151

Post by notahacker » Tue, Oct 10, 2023, 9:16 AM UTC

s3krit wrote: post #60

noelwelsh wrote: post #29

I'm sure this is impressive for an autonomous robot. However, as a fan of real parkour its kinda annoying to see some modest jumps and walking on a slope labelled "extreme parkour". What the robot demonstrates I'd expect any healthy 10 year old to be able to equal.

The metrics they're using (2x its height for climbing a wall, 2x its length for crossing a gap) are weird and don't really relate to the same achievements for a traceur. 2x its height is really more like slightly over 1x its usable body for that maneuver (0.4m length, 0.51cm height of the climb). I agree, not extreme but still pretty impressive for a robot. We're not going to see them doing cat leaps any time soon ;)

Relative size of leaps doesn't seem to be a particularly useful metric for assessing the NN performance anyway, since in the absence of the human constraint of fear it's really limited only by the mechanics of its legs and its weight.

More impressive would be adaptation to obstacles without clearly delineated edges, sticking landings on uneven/moving landing sites, and especially avoidance of landing sites which appear incapable of supporting its weight properly, particularly if it could do it well enough to generalise to novel courses.

The video hints the model may be able to do this to some extent (the high jump does show some apparently necessary compensating movement to avoid slipping off), but doesn't really demonstrate it.

Re: Extreme parkour with legged robots

#152

Post by Cyphonetic » Tue, Oct 10, 2023, 9:16 AM UTC

moffkalast wrote: post #143

TulliusCicero wrote: post #139

Earlier quoted context omitted.

Well, for one obvious reason, that's not terribly interesting from a scientific perspective. For the other, legs are more versatile than wheels, especially when it comes to uneven terrain.

The best solution would be to add wheels as feet which gives you the best of both worlds. It does however add even more cost to an already outrageously priced robot. https://spectrum.ieee.org/wheels-are-better-than-feet-for-le...

The robot that they used is 14k, down from boston Dynamics 75k, and they also have another which is available for 5k which you possibly might be able to do the same thing with it too.

Outragously expensive is not correct if you ask me.

Re: Extreme parkour with legged robots

#153

Post by michaelt » Tue, Oct 10, 2023, 9:25 AM UTC

Animats wrote: post #42

" A single neural net policy operating directly from a camera image, trained in simulation with large scale RL, can overcome imprecise sensing and actuation to output highly precise control behavior end-to-end. " So that's what it takes. That's so much simpler than the way Boston Dynamics does it, working out all the dynamics in simulation first. It's amazing to see this done from vision to actuators in one net. It's…

> "A single neural net policy operating directly from a camera image, trained in simulation with large scale RL, [...]

> So that's what it takes. That's so much simpler than the way Boston Dynamics does it, working out all the dynamics in simulation first.

I mean, it sounds like they have had to work out all the dynamics, and simulate them.

Admittedly they don't have to do it on-robot in real time any more.

Re: Extreme parkour with legged robots

#154

Post by cnity » Tue, Oct 10, 2023, 9:37 AM UTC

loandbehold wrote: post #105

lukevp wrote: post #79

It’s pretty incredible how animalistic its behaviors are becoming (hesitating for a moment at the edge of a lip, coiling it’s back legs for increased actuation, and when it almost misses the high jump and moves its back leg really fast multiple times to give itself a tiny boost each time until it recovers). Is that because it’s trained on animal behaviors? Or is this emergent behavior and animals just also do it more…

it suggests that animals learn using similar neural nets.

Or it simply suggests that this is advantageous behaviour and that different learning methodologies would result in the same behaviour.

Re: Extreme parkour with legged robots

#155

Post by developer93 » Tue, Oct 10, 2023, 9:39 AM UTC

quickthrower2 wrote: post #118

lukevp wrote: post #80

Earlier quoted context omitted.

The difference being each human has to spend thousands of hours learning that level of control over their bodies, and work hard to maintain that level of physical fitness. And if they fall off the roof, millions of dollars of potential earnings and GDP die with them. Whereas these robots are only $70k, and once one of them can do this, they all can do it. Just like with Chess and Go. It’s not impressive at first, the…

The most impressive thing about the video is not the skill but the control of their fear.

Not that have experience, but I feel like that's just practice. Your subconscious needs to learn that it's not necessarily going to die when it does that, and your subconscious learns from experience, so if you did it 10 times and nothing went wrong, you'll be less scared.

Re: Extreme parkour with legged robots

#156

Post by a96 » Tue, Oct 10, 2023, 9:46 AM UTC

foderking wrote: post #56

noelwelsh wrote: post #29

I'm sure this is impressive for an autonomous robot. However, as a fan of real parkour its kinda annoying to see some modest jumps and walking on a slope labelled "extreme parkour". What the robot demonstrates I'd expect any healthy 10 year old to be able to equal.

its actually hard for robots to compete with 10 year olds

But how amazing is it that this is even a thing to consider today

Re: Extreme parkour with legged robots

#157

Post by a96 » Tue, Oct 10, 2023, 9:53 AM UTC

riku_iki wrote: post #23

I am always wondering, why they build such complicated biology inspired robots, while they can build something with wheels and simple levels which allows robot to jump over obstacles and turn upside down if it landed on the back.

Somebody else is already doing simple wheels and levels. A lot of people/groups, actually. These and others are doing legs. I don't know if anyone is looking at crawling and jumping, or screws or tracks. Or tentacles. I wouldn't be surprised.

What would be really interesting if some way of motion turned out to be much better than what we find in nature or manufactured for non-autonomous machines.

Re: Extreme parkour with legged robots

#158

Post by a96 » Tue, Oct 10, 2023, 10:09 AM UTC

hanniabu wrote: post #43

stavros wrote: post #13

Earlier quoted context omitted.

Why is walking the salient technology that will enable this? Why not wheels, or treads, or flying?

It's not necessarily walking, just the improvement in mobility which is one precursor to self-sufficient policing

Or obstacle avoidance and hints of advancing capabilities for learning and adapting? (Apparent) autonomy in general?

Re: Extreme parkour with legged robots

#159

Post by waldarbeiter » Tue, Oct 10, 2023, 10:17 AM UTC

danmaz74 wrote: post #137

I scanned the paper and, if I got it right, during training the robot gets "external" information about its world position and speed compared to pre-set waypoints, something that an animal or person wouldn't have access to. What I didn't understand for sure is if this information is also available to the robot while performing the "parkour": I'm pretty sure that it perceives the obstacles only through its depth camer…

When the policy is deployed in the real world only the depth camera is used no waypoints etc.. Scan dots and target heading is used in the first Phase of the training to pretrain a policy in simulation. In Phase 2 a policy is trained end-to-end using the pretrained actor network: "First, exteroceptive information is only available in the form of depth images from a front-facing camera instead of scandots. Second, there is no expert to specify waypoints and target directions, these must be inferred from the visible terrain geometry." For policy training in Phase 2 DAgger which is based on Behavior Cloning is used (with the policy from Phase 1 as the expert), they also use some tricks to make sure no actions that are too different from the expert actions are executed during training. In Phase 2 the network learns to extract environment information instead of from the scan dots from the depth camera. Also in Phase 2 they use the pretrained actor network from Phase 1 but the depth embedding must be learned from scratch. This is how I understand it.

Re: Extreme parkour with legged robots

#160

Post by blagie » Tue, Oct 10, 2023, 10:36 AM UTC

beachy wrote: post #74

jiggawatts wrote: post #57

Earlier quoted context omitted.

Artillery is surprisingly inefficient. It takes about 15 rounds of 155mm artillery per battlefield casualty. Each shot costs about $5k. Robots like this might cost the same as a single artillery shot but have the ability to get multiple kills each. It’s a cold arithmetic, but I guarantee you someone at DARPA is thinking and planning the future of warfare in these terms.

I thought you must be wrong - $5K for a dumb arty round? But some googling indicates this is about right, and prices have even gone up since the Russian invasion.

I don't have real numbers for anything, but I tend to do analysis like the above quite often when evaluating business models. It's usually within a factor of 2x to 10x of being correct, so please don't take it as anything more than that:

A typical round weighs around 40 kilos or 100lbs. For reference, from a quick web search, as commodities:

- 100lbs of TNT would be around $500

- 100lbs of stainless steel is around $250

I'm not sure what materials go in (I'd assume better explosives and worse steel), but I think $250-$500 is a good estimate for price of raw commodities (not including anything beyond that).

I would at least double that for having those shaped into an artillery shell, and add the cost of non-commodities (like a detonator), and you're easily at $1000. I would at least double that for the raw cost of distribution, sales, shipping, and logistics, and other overhead. We're probably at $2000.

At that point, toss in 30% profit margin for everyone along the way (30% to manufacturer, 30% to distributor, etc.). You're probably now around $3500.

I think that's a bare minimum baseline for what it would take to get a very, very dumb shell. This goes up if you want:

- Anything at all fancy or high-tech

- War profits

- NREs covered (building a factory, which might be idle 95% of the time and spin up for wartime)

- Military inefficiencies

- Regulatory compliance, quality inspection, etc.

Etc.

You can hit $5k very easily.