Robust autonomy emerges from self-play | Hacker News Reader

Robust autonomy emerges from self-play

arxiv.org

140 pointsreqo62 comments

Some interesting points from this paper:

- All simulated agents use the same neural net with the same weights, albeit with randomized rewards and conditioning vector to allow them to behave as different types of vehicles with different types of aggressiveness. This is like driving in a world where everyone is different copies of you, but some of your copies are in rush while others are patient. This allows backprop to optimize for a sort of global utility across the entire population.

- There is no modeling of occlusion effects. Instead, agents are given the state of nearby agents, but corrupted by random noise. In the real world, occluded nearby agents can be extremely close (think about a child running out from behind a parked car). The paper comments on this.

> Both Waymax and nuPlan construct observations, maps, and other actors with auto-labeling tools from realworld perception data. This brings occlusion, incorrect or missing traffic-light states, and obstacles revealed at the last moment. Despite the minimalistic noise modeling in GIGAFLOW, the GIGAFLOW policy generalizes zero-shot to these conditions.

- The resulting policy simulates agents that are human-like, even though the system has never seen humans drive. This is a great result when one considers other reinforcement learning projects produce extremely high performance agents that humans would consider to be abusive or pathological.

Can there be "smart toys" for models that help them self-improve in a particularly efficient way?

Can you expand on that? Efficient in what way?

Efficient in the way of bringing the model to meet the criteria of autonomy faster. On one hand it may be something specifically efficient at reaching some autonomy qualities. OTOH it could be just something that efficiently uses the improvement in the model during training to make the subsequent training faster.

Yes, the smart toys are search, code execution, simulations and games.

video games are basically like this, progressive level require more skill, learned from the easier levels.

And this is a reason we play video games? That they appeal to some ancient instinct to improve?

One of the reasons. I'd wager this is what appeals to people not merely playing but mastering a particular game - playing higher difficulties, 100% completion, and so on.

The other reasons would be overcoming other humans (esports/pvp multiplayer), discovery (story driven and exploratory games), and just passing the time (casual games).

Rather, I think it’s borne of necessity to onboard you to the game mechanics. We certainly have a bird brained instinct to catch the worm / win a round, so a good game design cuts you some slack to begin with, so you can have a little dopamine as a treat

From there, difficulty should scale up so you don’t always win, giving you that “intermittent reinforcement“ that makes games addictive

I'd say it's feedback/reward loop plus small, quick to achieve, progressive goal setting.

In RL literature this is generally called "curriculum learning".

The curriculum is usually modeled as some form of reward function to steer learning, or sometimes by environment configuration (e.g. learn to walk on a normal surface before a slippery surface).

I'm creating "smart toys" like that for humans. I recently launched a mobile app. I'd love to see these research breakthroughs feed back into human learning because if humans remain foolish, the world could fall apart.

With DeepSeek R1 and these autonomous driving research results, it feels like we've entered an era where human data is no longer necessary. The ability to infinitely expand learning through simulation while maintaining safety in the real world feels like science fiction coming to life—it's truly exciting.

A Young Lady's Illustrated Primer

This is interesting, and I have always thought this approach worth exploring given the "bitter lesson" in other ML domains, but I think we should be skeptical until we see such models deployed and operating effectively on real-world vehicles.

Interesting to see this coming out of Apple

Something about dreams that fascinates me is that I usually am genuinely surprised by events that occur in dreams. I interact with other characters whose motivation I cannot understand and whose actions I cannot fully anticipate. It feels like there's a foreign entity acting as DM.

This isn't fake surprise. Sometimes I'll wake up and think, "who on earth were those guys and what were they trying to do? And yet their actions make sense..." or, "who came up with that punchline? It's legitimately funny and I never saw it coming, so it can't have been me..."

And yet I know it's all being generated by my own brain somehow. Through some kind of privileged access level.

And then I think about the bicameral brain structure. Does our brain have two halves so that it can function in a self-play training mode during sleep? Are each halves of my brain experiencing the same dream from opposite points of view?

Apologies for the tangent; this is almost totally unrelated to the article and probably something well known to neuroscience for decades. But still, it fascinates me, and the more we learn about the effectiveness of self-play in AI, the more I wonder.

I think you may have hit upon a novel combination of ideas here. There is something called "social simulation theory" regarding the purpose of dreams, but I don't think it has a neuroanatomical description included.

>"I don't think it has a neuroanatomical description included"

Genuinely curious, but then why bother?

because you can make descriptions of useful human cognition without specifying the implementation. When you’re reading python code, do you ask, “wait, which hard register is this local variable stored in??”

I don’t think this requires two halves, although it certainly seems possible that is what is happening.

I believe it only requires that your sensory and post-sensory systems be unpredictably generative when feeding to your subjective sense-making/observer. This could be provided for within a coherent whole brain.

That would work for dreams of wandering through some random landscape, but I'm talking about dreams where I interact with people who seem as smart (or smarter) than me, who can tell me jokes that make me laugh.

What you describe would be like learning chess by exploring random boards, but I'm talking about dreams as self play: learning chess by playing as white against black, without any window into black's strategy. To do that well seems to require running two brain instances in relative isolation. Dreams would be the only safe time to do that, and a bicameral brain hardware would be the most straightforward implementation. I doubt my optic nerve can play chess against my cerebellum.

Is it possible that you are confusing the feeling of the dream with the content of the dream? For example can you remember and write down any of these jokes? Do they make any sense outside of the dream?

The dream may give you the feeling that someone is telling you a joke, and the feeling that makes you laugh, without the actual joke existing as a real structured text joke.

I can, yes. I'm lucid enough after waking up to confirm that they're sometimes pretty good jokes, and very unlike what I'd come up with when awake. They're not comedy gold, I'm not a comedian, but they're but decent for me - and they would fall flat if you knew in advance how the setup was going to land.

I'm aware of the possibility you're hinting at; it's my default expectation for a dream to feel sensible but not be sensible. Like dreaming that I can speak Spanish fluently, but when I wake up I realize that none of those were real words. It's when this doesn't happen that I feel so amazed: when I wake up and reflect and it still feels like there must have been someone else in there with me, whose moves I was unable to introspect, but they were nonetheless coherent.

Both happens. Sometimes its only a feeling what covers the nonsense, other times I can remember the details pretty well (and its funny or extraordinary in other ways, like "meaning" has many layers, entirely unexpected).

I once had a lucid dream in which I attended a music concert, and the music was very beautiful, catchy, with complex professional arrangements; unlike anything I had ever heard before. Since it was a lucid dream, I felt proud that I could synthesize such cool music. But when I woke up, I couldn't remember anything to recreate it. This made me wonder, was the music truly that beautiful, or had I merely hallucinated my emotional response to it?

As I said, I’m open to the possibility that you are correct, but it isn’t accurate to reduce the piecewise analogy to an optic nerve playing chess against a cerebellum.

You have multiple visual cortices that are made of roughly the same stuff as the rest of your neocortex. There is more than enough idle network/processing capacity there, given it is not being fed visual stimulus by the optic nerve, to “play chess.”

It’s a very conscious part of sleep. So who knows what’s actually going on elsewhere, and if you ever make it conscious, would it just be an interpretation of this temporary conscious machinery inspecting what was previously running without it with no labels.

It is similar to solving problems. You want most of it to happen in unconsciousness, otherwise it’s too slow.

Things are learned when they are natural, without thought.

By the way, on dreams: Doesn't the existence of dreams mean that consciousness is like a subprogram in the brain that is usually "plugged in" to reality (or rather some rendering of reality done by the rest of the brain) but the rest of the brain can in fact simulate other scenarios for the consciousness to plug into?

Also I think split brain experiments basically support that you can split a consciousness into two tbh...

Anyways obviously above is speculation

That's one way to think about it I guess! I always think of it more like the channels that write to the sensory input registers get switched, just like when I configure my microcontroller's ADC0 to read from INT_REF instead of EXT_REF by flipping a bit at a certain memory address.

You know how a loud sound or a flash of light can alter your dream ... For example you hear your alarm clock and all of a sudden you are dreaming you are in an ambulance?

My theory is that we are always in a "dream" state. Stimuli that manage to reach our conscious attention will alter this dream.

When a asleep only very strong stimuli will reach us so for the most part our "dream" is in a feedback loop mostly doing its own thing. When awake though we have a much more weaker filter for stimuli . The direction that our "dream" takes is fully controlled by it.

> "who came up with that punchline? It's legitimately funny and I never saw it coming, so it can't have been me..."

Sometimes, when I dream, I am envious of people being witty in ways I think I can't be, and when I wake up I'm like "..."

Or dreams with music scores or dreams where you are composing a music score. But once I didn't quite wake up and I got it down in a sequencer. Freaked out my boss who didn't believe me when I said I made it.

I have this when I'm drifting off to sleep but not quite gone yet. I can effortlessly compose compelling melodies in my head, sometimes even parts for multiple instruments simultaneously, as if I'm simply listening to someone play it. But if I want it to go in a particular direction, it does. The couple of times I've willed myself to wake up and find a keyboard, the melodies still made sense — I wasn't just a drunk guy thinking he can sing.

I've occasionally been able to do the same with architecture: design a massive sprawling palace with ease as I fly through it. And much like the music, on the one and only (I think?) time I woke up and sketched as much as I could remember, it still all made sense.

But in my normal waking life I am creatively constipated. My mind aggressively criticises and crushes ideas before they get a chance to grow organically. I have one side of the creative process in my waking life (filtering) but very little of the other (synthesis).

This makes me think a couple of things:

1. I totally get why artists use drugs so much. Any way to tap into that other state must be incredibly tempting.

2. It would be so amazing if we could figure out how to record in high fidelity and interpret what's going on in these altered states of consciousness. Maybe you've composed a whole symphony while you slept once, and you just don't know it!

You should check out "the artist way" it's a book about releasing you creativity. One of the pillars is the concept of morning pages. Just a stream of thought writing 3 pages the first thing you do after you wake up. I guess it helps training yourself to be less in "critique mode"

>But in my normal waking life I am creatively constipated. My mind aggressively criticises and crushes ideas before they get a chance to grow organically. I have one side of the creative process in my waking life (filtering) but very little of the other (synthesis).

There's an exercise I've had a bit of success with over the years: Identify a problem, then write down the first ten solutions to it that come to mind, no matter what.

It's important to write down each response you think of, no matter how impractical, ineffective, implausible, or ridiculous. You'll start to notice when you hold back from writing something down, no matter if it's because you think it's unoriginal, it seems like it doesn't fit the problem, or if you think it's too taboo. It's that _noticing_ that you're training, that's the goal. All forms of preemptive filtering keep us from letting the beginning of an idea take root, so the other ideas that might have followed from it never even occur to us.

Children are much better at this game than adults. When asked to invent a better paper clip, adults tend to focus on how to improve a standard paper clip's paper-clipping, how economically it could be made, or to focus on and change one feature at a time. Children can come up with ideas like building-sized paper clips made of cotton candy, to suggest a pony could be a good paper clip because you could tell it when to open or close its mouth and because it makes them happy, or to tell you how faeries would make paper clips so they could fly around on them.

It can help to have a good way to avoid feeling too embarrassed to write down every idea. Maybe you feel less inhibited when you pre-commit to deleting everything you typed, or tearing up and burning the paper you wrote them on after you're done. I know that saving the ideas makes me more likely to judge them by how other people might react, but perhaps saving the ideas might help for others.

Don’t remember where I read this theory, but the gist is that dreams are simulated events for practice/preparation, and they are therefore surprising and very diverse - you need to practice social interactions, fight or flight, shame, freezing, etc - and since you don’t have a lot of time, many dreams will sample many or all these themes.

I'll be grateful for all the prep dreams when I finally find myself doing an exam I'd forgotten about while naked.

I tend to rely on old traditional methods and recently came across this dissection which opened up my interpretation of the whole dreaming process (a way in which we are "renewed"): https://www.youtube.com/watch?v=uz4SRe-PH7A

The problem is, you might dream the joke is very witty, but when you wake up, you realize it is completely nonsensical.

Your brain didn’t really dream the joke. It dreamt the emotional response to it. The joke itself was just a prop.

Not sure where I’m going with this.

Also happens with drugs, the you understood everything thing, teleportation, repetitions, etc, repetitions, etc.

We get used to our usual normal-functioning consciousness, but there’s a whole universe of its potential modes and what-causes-what and in which order. For example, dream scenarios may easily happen in reverse, from a random emotion to drawing the scene around that, to the sense of reality, when you just remember a concept of it. Everything can be backwards, orthogonal, not in order. All “it” has to do is to make it feel normal. All these “backwards”, “normal”, “time”, “because” are not something just granted, you have to actively experience these all the time.

In that case, it wouldn't still be funny and unexpected when I woke up.

> I interact with other characters whose motivation I cannot understand and whose actions I cannot fully anticipate. It feels like there's a foreign entity acting as DM.

My head canon on this phenomenon is that we are not quite as integrated (or isolated) as our conscious 'self' would have us believe.

Neuroscience and Buddhism alike seem to back this up... The concept of 'anatta' [0] has held up strongly for thousands of years (how could it not?). As far as I can tell, neuroscience is also increasingly clear that consciousness is an emergent phenomenon - and strange things happen when this is disrupted.

Most who have experienced psychedelics will agree that our brains are capable of providing utterly novel, previously unimaginable experiences; stranger even than most dreams.

For example, DMT users across a wide range of cultural backgrounds describe seeing 'machine elves' [1]; "People describe these entities as distinct, autonomous beings that typically present with some kind of message."

Then, look at people who 'hear voices'; a phenomenon which is quite scary in our culture, and quite normal or respected in other, older cultures. The voices can say things which surprise us even in our waking lives, sometimes even offering powerful insight, without any ingestion of substances.

So... There are these ways to experience surprising characters made by our own brains - dreams, drugs, and other oddities. My personal view is that these are always active, mostly doing their thing in the background; something like flora and fauna in the sea of our sub/consciousness. Sleep can reveal them, as can 'altered' states of mind.. As if the light of consciousness dazzles us, and we can't see the gloriously intricate machinery in the dark.

0 - https://buddhism.stackexchange.com/a/25302

1 - https://health.howstuffworks.com/wellness/drugs-alcohol/dmt-...

What I always wonder when this happens is, did I actually come up with something witty in a dream or did my brain just generate the feeling of having come up with something witty without actually having a cause.

I had a dream once where I had 3 really good ideas for a particular product but when I woke up I couldn't remember the 3rd and best idea. Then I started wondering if there actually was a 3rd idea or did my dream just skip to the "That's a great idea" stage with no actual idea, leaving me with just the feeling.

Brains are weird, dreams doubly so.

"Why would you need to see it? You're the genius who invented the product in question!"

> https://www.youtube.com/watch?v=9LtT0xZ11wM

Brilliant clip. :) I'm sure we've all had dreams that feel like that. If I had to put a point on it for this one, the "self-play" game happening here is your brain serving up instantaneous new plausible and unanticipated excuses for why you can't see the object that you're struggling to get a glimpse of!

Maybe not directly related , I find genertic algorithms and other optimisation algorithms such as Ant Colony Optimisation algorithms intersecting with this approach of self-play and leading to robust autonomy.

The concept of being able to simulate 42 years of “experience” in one hour seems so foreign to me. Something about it creeps me out.

Don't watch the White Christmas Black Mirror episode.

Maybe at this point don’t watch any black mirror episodes…?

This is also an acceptable option. It's good TV but it's also nightmare fuel.

Don't read Junji Ito's Nagai Yume either.

Humanity experiences almost a million years per hour.

If the gym is sufficiently simple and well-coded, achieving a simulation speed of 367,920x real-time (simulating 42 years in one hour) is plausible. The question is whether these simulated scenarios genuinely reflect 42 years of real-world driving experience and truly represent the information that a single agent has at its disposal when making driving decisions.

I had a couple of hobbies (lucid dreaming and shamanic trance drumming) that enabled me to experience big disconnects between the subjective experience of time passing and objective measurable wall-clock time. Some dreams and trances subjectively appeared to be much longer than the wall-clock time recorded by clocks and human helpers.

I don't have any definite knowledge of what's going on with that, but I suspect some part of it is my brain retroactively manufacturing the memory of lots of time passing, and some part of it is my brain confabulating episodic memory about the dream or trance as I wake up and write it down.

Human memory is well known to be generally unreliable and full of confabulated details, so I think the most parsimonious explanation for differences between the time experienced in dreams and the objectively-measurable time that passes is that our brains are just making shit up.

Of course, the idea that your brain just lies to you about the past might be just as creepy as any other explanation.

[stub for offtopicness]

[flagged]

"Guys, do you think we'll get away it?"

"Self-play simulation emits fitter models"

"Self-play has powered breakthroughs in two-player and multi-player games."

[flagged]

Could you please stop posting unsubstantive comments and flamebait? You've unfortunately been doing it repeatedly. It's not what this site is for, and destroys what it is for.

If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.

Reply on news.ycombinator.com