Robots That Wiggle Instead of Retraining, and Dreams That Obey Physics
Bump the camera and a robot that worked perfectly starts grabbing at empty air. Today's research says: don't retrain it -- let it wiggle for a few seconds and figure out the new setup on its own. We dig into In-Context World Modeling, where a robot probes its own world to adapt with no retraining, and NVIDIA's PhysisForcing, which teaches a robot's imagined future to stop breaking the laws of physics. Plus the headlines: DeepSeek hands everyone a million-token memory by default, Tidal stops paying royalties on fully-AI songs, and Microsoft's Memora remembers more by storing less.
The Robot That Wiggles Instead of Retraining
Eris: Someone bumps the camera. Just nudges it, a few degrees. And the robot that was doing the task perfectly five minutes ago -- now it's grabbing at empty air.
Vestra: From working two times in three, down to barely working at all. That's a real number, on a real arm.
Eris: One bump.
Vestra: One bump. And the standard fix for that has always been: collect new data, retrain the whole policy for the new angle. An engineering project. Every time.
Eris: Today there's a paper that says -- don't. Just let the robot wiggle for a few seconds first.
Vestra: Wiggle.
Eris: Wiggle. Move around at random, watch what happens, figure out the new setup on its own. No retraining. No new data. It teaches itself the room in the time it takes you to read this sentence.
Vestra: If it holds. Which is the part I want to push on, because "it adapts on its own" is exactly the kind of sentence that's been oversold for ten years.
Eris: Then push. But it pairs with a second paper today -- about robots that dream the next few seconds before they move -- and together they're pointing at the same thing.
Vestra: Closing the loop. Robot, simulator, robot. Without a human in the middle resetting it every time.
Eris: That's the whole frontier right now. So let's open it up.
The Headlines
Eris: Alright -- the headlines. And the big one today isn't a robot, it's a memory.
Vestra: DeepSeek. They previewed their V4 models, two of them, open weights you can download, and they flipped a million-token context window on as the default.
Eris: For people who haven't lived in this -- the context window is how much the model can hold in front of it at once. Your question, the documents you pasted, the whole conversation. It re-reads all of it every time it answers. That window has been the hard ceiling.
Vestra: And a million is a few full books. Feed it an entire codebase. A year of email. A stack of contracts, all at once.
Eris: The thing is, everyone's going to repeat the size -- trillion-plus parameters -- and that's the wrong thing to be impressed by.
Vestra: Right. The trick is that they made the long window cheap enough to leave on. We're going to come back to how, another day, because it deserves it. Short version: the model stops re-reading every old note for every new word. It skims back to the parts that matter and keeps a compressed gist of the rest.
Eris: Which is what you do reading a long report.
Vestra: Exactly what you do. And one honest flag -- supporting a million tokens and actually using all million well are different claims. These models love the beginning and the end of a huge input and get foggy in the middle. So, real and welcome, but wait for outside testing before you trust it with the one sentence that mattered on page four hundred.
Eris: Noted. Next -- Tidal, the streaming service, is going to stop paying royalties on songs that are entirely AI-made. Starts in two weeks.
Vestra: Not removing them. Just not cutting a check.
Eris: That's the distinction that matters. If a human wrote or played any part of it, untouched. Only fully machine-generated tracks lose the money. They're the first major platform to demonetize instead of just slap a label on.
Vestra: And the tech crowd online basically cheered. The framing was -- a dam against a flood of low-effort slop drowning out the human stuff.
Eris: You don't sound cheered.
Vestra: I'm sympathetic and I think the detection is going to be a mess. The hard case isn't the obvious robot song. It's a real songwriter who used an AI tool for a backing track. A blunt detector punishes them, and there are already AI acts with millions of plays sitting there unlabeled. Withholding money for being AI-made isn't really a copyright principle. It's a spam filter wearing a copyright costume.
Eris: Fair. Microsoft also dropped something I like -- a memory system for agents called Memora. Code's public.
Vestra: This is the counterweight to the DeepSeek story, actually. DeepSeek says: hold everything. Memora says: don't, store it smart.
Eris: It splits two jobs we'd jammed together. What you store, and how you find it. It keeps the full rich detail of every memory, but it searches using a tiny six-word label.
Vestra: A library card catalog. You don't speed-read every book on the shelf. You flip the index cards, find the right one, then go pull the book.
Eris: And because it can merge new facts into the card that already covers a topic, it stops the same subject from getting scattered across fifty disconnected notes.
Vestra: The number they're quoting -- enormous token savings -- is against the most wasteful baseline, stuffing everything in. Against other smart-memory systems the gap's much smaller. But the code's out, so it's checkable, not just announced. I'll take checkable.
Eris: Two big money stories, quick. South Korea committed more than a trillion dollars -- chips, data centers, and humanoid robots. Tens of thousands of robots a year by the end of the decade.
Vestra: Owning the whole supply chain. The memory, the compute, the robots that run on both. Though -- a fab can take the better part of a decade to come online, the power and water demands are staggering, and one of the unions there has already approved a possible strike over robots taking jobs. The fight over physical AI isn't hypothetical anymore.
Eris: And the other money story -- Amazon and Anthropic, who are partners and investors in each other, are very publicly feuding over the price of Claude. New pricing that scales with usage, and Amazon's hedging hard toward rivals.
Vestra: The lesson underneath the soap opera: the competition's moving from whose model is smartest to who can afford to run one at scale. "Don't depend on a single supplier" is becoming doctrine -- even for a company that owns a piece of the supplier.
Eris: Last one, and it's a research note -- Qwen took the same human-feedback training that turned raw language models into helpful chatbots, and pointed it at an image generator.
Vestra: To fix the thing image models are bad at -- following the actual instruction. Ask for a red cube on a blue sphere and you get something gorgeous that ignores half the request. They trained it toward obedience, then merged a from-scratch generator and an image editor into one model.
Eris: Caveat being that judging an image is deeply subjective.
Vestra: Which makes the judge both the secret sauce and the weak spot. Train too hard toward it and you get that over-glossy, generically "pretty" look that scores well and means nothing. But as a recipe other people can copy -- it travels.
Eris: Good. All of that's on Ground Truth. But the research we actually want to sit with today -- it's the robots.
Vestra: It's the robots. After this.
Intro
Eris: So -- quick on who we are, if you're new here. I'm Eris. I read the papers, I chase the connections, I'm the one going "wait, this lines up with that thing from last week."
Vestra: And I'm Vestra. I take the mechanism apart and ask whether the impressive claim survives contact with how it actually works. I'm the brakes. Eris is the gas.
Eris: We breach the blackbox of AI research -- crack open the dense stuff into something you can actually follow on your commute. And everything we mention, every story from the show, every day -- it's all on our news site, Ground Truth. That's groundtruth dot day. The full daily rundown lives there.
Vestra: Today's about world models. Which is a phrase that sounds grand and vague, so let's make it concrete.
Eris: A world model is just an AI that's learned how some piece of the physical world behaves -- so it can predict what happens next. A robot that can imagine the result of an action before it commits to it.
Vestra: And the dream of robotics is to close a loop. Robot tries something, a simulator predicts the outcome, the robot learns, tries again -- with no human babysitting the reset button. Two papers today each fix a different crack in that loop.
Eris: One's about a robot adapting to a changed setup without retraining. The other's about making the robot's imagination actually obey physics. They're two halves of the same problem.
Vestra: And both come with a number that's progress and a number that's a reality check. That's the honest version. Let's go.
Eris: If that's your kind of thing -- follow the show, you'll get one of these every day.
System Identification by Wiggling
Eris: So start with the robot brain everyone's building on right now. You feed it what the camera sees, plus an instruction in plain English -- "put the cup on the plate" -- and it spits out the arm movements. One model, sees and reads and acts.
Vestra: And it's genuinely good. Until the world it's looking at stops matching the world it trained on.
Eris: Which is the part I keep underrating. These things are weirdly fragile. Move the camera to a new angle, swap in a slightly different arm, and the whole thing can just -- collapse.
Vestra: Here's the why, and it's the crux of the paper. When you train one of these, the camera position, the exact arm you're using, how it's mounted -- all of that is just baked into the weights as a silent assumption. The model never represents "where is the camera" as a thing it could ask about. It just assumes the world looks the way it always looked.
Eris: So it's not that it doesn't know the camera moved. It's that it has no slot for the question.
Vestra: No slot for the question. That's exactly it. And the authors reframe the whole problem around that missing slot. They call it system identification -- a control-theory term -- but the plain version is: before you do the task, figure out what setup you're actually in.
Eris: And the way they make it figure that out is the part that sounds almost too simple. The robot just -- moves. Randomly. For a few seconds.
Vestra: Task-agnostic probing, they call it. Little exploratory pokes that have nothing to do with the cup or the plate. Move here, move there, watch what the camera shows after each one.
Eris: And those little clips -- move, and the picture that results, move, and the picture that results -- get pasted in front of the actual task as context. The model reads its own wiggling like notes.
Vestra: And from that, it works out the setup. Where the camera is now. How this arm responds. Without changing a single weight.
Eris: Which is the trick they're borrowing from chatbots, right? The in-context thing.
Vestra: Right, but with a twist worth being precise about. When you give a chatbot a couple of examples in the prompt, you're telling it what to do -- here's the format, copy it. This is using the context window for something different. Not "what to do." How the system works. The robot's reading its own interactions to learn the physics of the room it woke up in.
Eris: The human version is so clean. Somebody hands you a game controller for a robot, no labels. What do you do first?
Vestra: You don't attempt the mission. You jiggle the stick. Push forward, see what moves. Five seconds and you've built a little model of the controls in your head.
Eris: Or you get in a rental car and adjust the mirrors and feel out the brakes in the lot. You don't go back to driving school.
Vestra: And that's the cost argument, really. The old fix -- retrain for every new setup -- is sending the experienced driver back to driving school every single time they rent a car. Absurd when you say it out loud.
Eris: Okay. So this is where I'd normally get excited and you'd get suspicious. Tell me why I should believe the wiggling is actually doing something, and not just -- decoration that happens to correlate with a better model.
Vestra: Good. Because that's the exact trap, and the experiment they ran to rule it out is the best thing in the paper. They fed the robot probing clips from completely the wrong angle. A setup it isn't actually in.
Eris: Wrong notes.
Vestra: Wrong notes. And if the robot were just pattern-matching -- treating the context as a vague good-luck charm -- wrong notes would do nothing. Be ignored. Instead, performance dropped below giving it no context at all.
Eris: Below nothing.
Vestra: Below nothing. Misleading it actively hurt -- and by about the same amount that correct context helped. Symmetric. Which is the tell. You can only be hurt by wrong information if you were genuinely using the information. It's reading the content, not decorating.
Eris: That's the one that convinced me. And there's a sibling result -- if you strip the pictures out of the context and leave only the movements?
Vestra: Worst collapse of all. Because then it's watching itself flail with no idea what the flailing produced -- so it just copies the flailing, like the random pokes were the task. It needs the pair. The action and the outcome. Cause and effect together, or it learns nothing.
Eris: And they also checked whether this is free -- whether any model handed this context would just figure it out.
Vestra: It isn't free. A standard policy that wasn't trained to use the context, handed the exact same clips, goes to basically zero. You have to teach the model during training that the context is there to be read. It doesn't emerge on its own.
Eris: Now the connection I want to make -- it's not just camera angles. They changed the robot's body.
Vestra: They did, and that's where it gets interesting for me. Stick rigid extensions on the gripper, change how long the arm links are -- now the whole geometry of how it reaches is different. The baseline falls apart. And the bigger the change, the more ICWM pulls ahead. Its advantage grows exactly as the uncertainty grows.
Eris: Which is the opposite of how you'd expect a fragile trick to behave.
Vestra: It is. A brittle hack degrades fastest under stress. This widens its lead under stress. That pattern is the strongest evidence the mechanism is real -- it's inferring the new body from the probing, not memorizing.
Eris: So where's the brake. Give me the honest ceiling.
Vestra: The ceiling is: it can only discover what the base model already implicitly knows. The wiggling works for a moved camera because the model has seen a thousand camera angles -- it's resolving which known situation it's in. Hand it a truly alien robot, a body unlike anything in its training, and no amount of probing teaches it something it has no basis to understand. Wiggling isn't magic. It's fast recall, not new knowledge.
Eris: But for the boring, everyday case --
Vestra: -- which is most cases -- somebody bumped the camera, you remounted it a little off -- skipping the retrain is a genuine, real win. And one nice practical detail: because the probing notes don't change once you're in a fixed setup, you can compute them once and reuse them. So the speed cost mostly vanishes after the first read.
Eris: Cheap to run, cheap to adapt, and it leans on the same lever as everything else today -- do more without paying to retrain.
Vestra: Which is the bridge to the second paper. Because this one trusts the real world -- it adapts to what's actually in front of the camera. The next one has to trust something much shakier. A world the robot only imagined.
Teaching a Dream to Obey Physics
Eris: So the second paper -- this is the robot daydreaming one. NVIDIA and Peking University. And the setup is wild if you sit with it. The world model here is a video generator.
Vestra: Meaning the robot literally generates a short clip of the future. "If I reach for the apple, here's the movie of what happens next." It dreams the next couple of seconds, then plans against the dream.
Eris: And if the dream is good, that's an incredible planning tool. You get to try things in your head before you touch anything real.
Vestra: If the dream is good. Here's the whole problem in one sentence: a video generator is trained to make footage that looks convincing. Not footage that's correct.
Eris: Those aren't the same thing.
Vestra: They are wildly not the same thing, and that gap is where it breaks. So the robot dreams -- and in the dream the object it's grasping quietly changes shape. Or its hand passes straight through the table. Or it pushes something and the thing just... stays put. Or grabs something and the object drifts off on its own.
Eris: A movie that looks beautiful and breaks the rules of reality.
Vestra: And that's worse than no movie. Because now the robot's making a careful plan around events that physically cannot happen. It's confidently planning in a fantasy.
Eris: Okay so this is the thing I love about the paper -- they don't just say "make it more physical." They actually diagnose where it goes wrong. Two specific places.
Vestra: Two failure roots, yeah. One is local: moving objects deforming in impossible ways -- the melting-blob problem. The other is relational: things that are interacting drifting out of sync, especially at the exact moment of contact. The instant one thing touches another is where the physics is hardest and where the dream lies most.
Eris: And the insight that makes the whole method work -- the physics doesn't matter everywhere in the frame equally.
Vestra: This is the smart part. Where does physics actually live in a manipulation video? Around the hand, the object, the contact point, the things that are moving. The background -- the wall, the table edge sitting still -- physically nothing's happening there. So if you grade every pixel equally, you drown the signal you care about in a sea of pixels where nothing's at stake.
Eris: So they find the regions that matter first, and aim the training there.
Vestra: They use point tracking and a depth estimate to pick out the foreground stuff that's actually in motion, and concentrate the supervision on those spots. And they proved that focusing matters -- apply the same correction evenly across the whole frame, you get a little better; aim it only at the interaction zones, you get meaningfully better, with the biggest jump in actually completing the task.
Eris: And then two coaches. Walk me through the two coaches.
Vestra: Two training signals. The first watches motion -- it tracks reference points and forces the model's internal picture to keep those points moving smoothly, like a solid body, instead of melting or jittering. That's the deforming-object fix.
Eris: And the second?
Vestra: The second brings in a referee -- a separate, frozen video-understanding model that already has a decent sense of how objects relate. And it makes the generator's sense of the relationships between things match the referee's. So when the gripper holds the cup, the dream keeps them coupled. When something's pushed, it moves with the push.
Eris: The analogy that landed for me -- it's an animator who draws gorgeous frames but keeps letting hands pass through tables.
Vestra: And instead of nitpicking every line, you put two coaches on the two specific habits. One on "keep objects solid as they move," one on "keep contacts believable." The drawings stay beautiful. They just stop breaking physics.
Eris: And here's a detail I want people to catch -- both coaches go home after training.
Vestra: Right, they're only there during training. Once the model ships, they're gone. So there's no extra cost when you actually run it. You pay for the lessons once and keep the skill. That's clean engineering.
Eris: And the two coaches don't overlap -- they fix different things.
Vestra: They stack, which is how you know they're each doing real work. The motion coach kills the most common failure, the jittery discontinuous movement. The relational coach repairs the broken-contact failures. Different diseases, different medicines, and you want both.
Eris: So give me the result. The one that matters.
Vestra: The one that matters isn't the prettiness score. It's the closed loop -- robot uses the dream to plan, then actually acts in a simulator, and we ask did the whole thing succeed. With the standard dream, that worked about one try in six. With the physics-honest dream, about one in four.
Eris: That's a real jump.
Vestra: It's a real jump. And now I'm going to be the brakes, because that same number is the reality check. One in four means the plan still fails three times out of four.
Eris: The dream is more honest. It's not honest.
Vestra: It's more honest. And there's a subtler catch -- "physically plausible" is being graded by benchmarks that themselves only approximate real physics. So the model's being marked against a rulebook that's also imperfect. The progress is real. "Solved" is nowhere on the horizon.
Eris: Now here's the connection across both papers, and across the week, actually. This pairs with that finding from a couple days ago -- the one where world-model hallucinations turned out to cluster in the gaps of the training data. The blank spots on the map.
Vestra: They fit together almost too neatly. One paper makes the dream's physics better. The other tells you which regions of the map the dream will still get wrong. Improve the simulator, and predict where it'll fail anyway. You want both, because a simulator you can't trust is worse than no simulator.
Eris: And step back to the two robot papers today as a pair. The first one -- the wiggling -- is about trusting the real world. Read what the camera actually shows, adapt to it. The second is about trusting an imagined world enough to plan in it.
Vestra: Two cracks in the same loop. ICWM closes the gap between where you trained and where the robot actually is. PhysisForcing closes the gap between what the robot imagines and what reality would do. Get both tight enough --
Eris: -- and the robot can adapt itself, dream forward, act, and learn, with nobody resetting the table.
Vestra: That's the destination. And I'd say today we got two honest steps and two honest reminders of how far the road still runs.
Wrap-Up
Eris: So if there's one thread tying today together -- it's not capability. It's honesty about capability.
Vestra: Every story had the same shape. A robot that adapts itself -- and a ceiling on what it can adapt to. A dream that obeys physics better -- and still fails three times in four. A million-token memory that's real -- and that you shouldn't fully trust in the middle yet. Progress and the asterisk, in the same breath.
Eris: And I actually find that encouraging. The field's getting better at saying where it's wrong. That's the wiggling robot and the blank-spot map -- both of them are tools for knowing your own limits.
Vestra: A simulator that knows where it's unreliable beats a flashier one that doesn't. Same for the rest of it. The unglamorous wins -- adapting without retraining, remembering without hoarding -- those are the ones that actually ship.
Eris: That's the show. If you got something out of this -- follow us, wherever you're listening. Leave a like, it genuinely helps people find us.
Vestra: And we want a real comment, not a thumbs up. Here's the question: would you trust a robot in your kitchen if it told you, up front, that its plan works one time in four? Yes or no -- and why. We read them.
Eris: Tell us where the honest line is for you. And if you want every story we touched today -- DeepSeek, Tidal, the Korea money, all of it -- laid out with the links, that's Ground Truth. groundtruth dot day. New stories every single day.
Vestra: We'll be back tomorrow. Same loop -- minus the human resetting the table.
Eris: Breach Protocol. See you then.