The robot intelligence gap

Two papers and two funding rounds landed in the same week, and together they map where embodied AI stands in late 2026.

On the research side, a team showed that vision-language-action models (VLAs) can fail on a synonym. Another team released a world model that finally turns video prediction into zero-shot manipulation. On the market side, a company that has never sold a robot is worth $10 billion, and a ten-person startup with no product is worth $200 million.

The common thread: everyone is betting that the software layer is where robot value concentrates. The research says that software is more brittle than its builders hoped and more fixable than its critics assumed.

One word can cost you 98 points

π0.5 turns on a LIBERO stove 100% of the time when the instruction says "switch on the stove". Change it to "switch on the hot plate" and the success rate collapses to 2%. Same scene, same physical action, one phrase swap, a 98-point swing. For a deployed robot, that's the difference between an operator typing the right synonym and watching the arm stand still.

That finding comes from "Rephrase Before You Act", a characterization of language sensitivity in VLAs. The authors ran statistically tested single-edit experiments and found swings everywhere. A π0 checkpoint finetuned with rephrase augmentation, which should be the robust one, still swings up to 61 points from one word.

The core finding: VLAs do not inherit the language robustness of the vision-language models they're built on. The base VLM understands paraphrases fine. The VLA, trained with action heads on top, throws that away.

The most useful number is the 21-point gap between in-distribution and out-of-distribution task success. An oracle phrase search nearly closes it. The model has the knowledge. The failure is in decoding intent from wording. That's good news: the knowledge is there, and the problem sits in a layer you can fix without retraining.

Fix the phrasing, don't touch the weights

The fix is unglamorous, which is why I like it. Because the sensitivity is systematic, you can express it as rules. The pipeline scores many phrasings of a few training tasks, has an LLM distill the evidence into ten to twenty rephrasing rules, and rewrites each incoming instruction once at deployment. No retraining. No per-step verification. No changes to the frozen policy.

The rules improve frozen π0 by 16 to 27% relative on twelve held-out tasks, across adversarial, VLM-generated, and human phrasings. Gains concentrate on out-of-distribution tasks, which is where a generalist robot needs help most. The same pipeline lifts π0.5's LIBERO in-finetune success from 93.6% to 97.8%.

Ten to twenty rules is small enough to audit. You can read exactly what the robot is being told to do differently. And because the rewrite happens outside the model, you can swap policies without re-deriving the rules. It applies zero-shot to unseen tasks and instructions.

Quick Take: VLAs are shippable today if phrasing is treated as part of the deployment stack, because the policy is mostly fine and the wording channel is the actual weak link.

World models that act: VPP2

Video Prediction Policy 2 attacks a different failure. World action models (WAMs) transfer video prediction priors to action learning, but existing ones produce wrong motion predictions in open-ended environments. The authors blame two things: base video models aren't optimized for manipulation, and naively adding action components degrades generalization.

VPP2's recipe has three parts. First, continue pretraining the base video model on a large, diverse manipulation video dataset with detailed captions and event-level pretraining. Second, distill it into a single-step visual planner with a fixed prediction horizon. Third, add an action module as a mixture-of-transformers (MoT) that learns an implicit inverse dynamics model: given a desired future, what actions produce it.

Results: VPP2-14B beats Cosmos3-64B by 11 points on video prediction instruction-following in open-ended tasks. It beats the strongest baseline by 18.5 points on real-world zero-shot ALOHA manipulation. After benchmark-specific post-training, it tops LIBERO-Pro, LIBERO-OOD, and RoboDojo.

Read the numbers with care. 14B parameters means cluster-scale inference, not something on a robot's onboard computer. The ALOHA result is the generalist claim: 18.5 points of zero-shot improvement on tasks the model was never finetuned for. The LIBERO results come after benchmark-specific post-training, so don't quote them as zero-shot.

The MoT design is the part worth stealing. A pure video world model predicts what happens next but can't tell you what to do. Separating visual planning from inverse dynamics lets each module specialize. That's the same decomposition instinct as the rephrasing rules: don't ask one component to do everything.

When 44,000 hours of data buys 0.5 dB

Nvidia's DreamDojo world model got an ICML spotlight. The recipe: collect roughly 44,000 hours of human video, add a few hundred hours of robot data, and pretrain a robotics foundation model starting from Cosmos 2.5. Table 4 shows about 0.5 dB PSNR improvement over the starting point.

Stop on that number. 44,000 hours is about five years of continuous footage. 256 H100s is a multimillion-dollar training run. 0.5 dB PSNR is the kind of gain you'd expect from a minor codec tweak. The resource-to-result ratio should have rung every reviewer's alarm.

I pulled the released repo over a weekend. The strange part: I could reproduce their numbers. Strange because a colleague ran the post-training script through Claude and found a logic bug that invalidates the post-training path. Then the GitHub issues surfaced two more bugs, one of which hits pretraining and evaluation. Three phases, three bugs. The paper's marginal results suddenly made sense: a broken pipeline produces a broken baseline.

What the community is saying, in my read of the thread, is less about Nvidia specifically and more about the system. The authors are well known. The paper got a spotlight. The code was public. Nobody asked why 44,000 hours of data bought 0.5 dB. The incentives reward a flashy but marginal foundation-model paper over the harder question: what did that data actually teach the model?

The real lesson: in robotics AI, a benchmark is only as trustworthy as the pipeline that produced it. Reproduce before you cite.

Cheaper legs: hierarchical control for quadrupeds

Energy efficiency doesn't make headlines, but it decides whether a field quadruped runs for forty minutes or two hours. End-to-end RL policies couple gait selection, motion execution, and energy optimization into one reward, and that coupling makes them hypersensitive to reward design.

This paper's fix is hierarchical RL. A high-frequency policy handles stable joint-level motion execution. A low-frequency policy adapts the gait and explicitly minimizes cost of transport (CoT). A three-stage Isaac-based training procedure enables zero-shot sim-to-real: the learned hierarchy was deployed on a physical Unitree AlienGo with no real-world finetuning.

The hierarchy produces automatic speed-dependent gaits, pacing at low speeds and trotting as commanded velocity rises. That's the gait pattern you'd expect from biology, and it emerges without hand-switching. Against single-policy and hierarchical baselines, the approach cuts CoT across a broad range of commanded velocities while keeping the robot stable on flat, rough, and inclined terrain.

Lower CoT means longer missions and less heat. Zero-shot transfer means you skip the painful real-world RL stage. Both matter more than another point on a manipulation benchmark.

The $10B brain with $135M in revenue

Now the market. FieldAI sells a "general-purpose robot brain": one software stack that drives humanoids, drones, and industrial arms. It sells no bodies, no joints, no hands. This week it's reported to be raising $700 million at a $10 billion valuation, five times its mark a year ago, with Hyundai among the investors. Cumulative revenue and contracts since founding: $135 million.

$10 billion divided by $135 million is about 74x. For a company whose only revenue stream is software licensing on hardware that hasn't mass-scaled, that multiple buys an option on the terminal state.

Company	Product	Valuation	Revenue or shipments	Sells hardware?
FieldAI	General-purpose robot brain software	$10B	$135M cumulative revenue and contracts	No
Zibian	Robot brain, deployed in homes and factories	~200亿 as reported	Not disclosed	No
Skild AI	Generalist robot brain	Not disclosed	$100M annual recurring revenue	No
Yinkesi	Robot joints, dexterous hands	A few hundred million RMB raised	100,000 units shipped; BYD and Zoomlion among customers	Yes

Key numbers: $10B valuation, up 5x in a year. $135M cumulative revenue and contracts. A 74x valuation-to-revenue multiple. $100M ARR for Skild AI, the most real of the brain startups. The parts makers, by contrast, ship 100,000 units to actual manufacturers.

FieldAI's pitch is the Android analogy: one brain, many bodies, like one OS, many phones. The problem is the installed base. Android amortized R&D across billions of handsets. The humanoid industry just celebrated a 5,000-unit order from Tesla. 5,000 units is less than a day of phone shipments. You can't amortize a $10 billion software bet on that.

The uncomfortable question is whether the analogy holds at all. Google won the OS war partly because it didn't build phones and opened the system to everyone. Tesla and BYD are building their own brains and their own systems. If the biggest robot builders keep intelligence in-house, the addressable market for an independent brain company is far below the OS fantasy.

You can read this two ways. The bubble read: 74x revenue with no installed base is a correction waiting to happen. The prelude read: if robots reach phone-scale ubiquity in five years, today's marks look cheap. The gap between those readings is a timing question, and nobody has real data on it.

World models, priced before they exist

Tian Keyu raised close to $30 million at a $200 million post-money valuation for a company with no public name, no released model, no product, and about ten people. Wuyuan Capital and IDG participated. The plan: release a complete world model around 2027.

The collateral is a NeurIPS 2024 best paper. Tian's Visual Autoregressive Modeling (VAR) generates images scale by scale rather than pixel by pixel or token by token, which speeds up generation. The startup extends that idea: a visual symbol system of roughly 200,000 symbols, trained on 100 million hours of video, aiming to cut video generation cost by at least an order of magnitude.

Context check. AMD announced an $8.2 billion all-stock acquisition of Fei-Fei Li's World Labs in September. DeepMind's Genie 3 generates real-time explorable environments at 24 frames per second and holds consistency for minutes. PixVerse's R1 and R2 push continuous interactive worlds. World models are the most crowded frontier in AI, and the field has no standard benchmark. That's why investors price papers and people instead of products.

The backstory adds friction. ByteDance dismissed Tian in 2024 for interfering with another intern's training job, and sued for 8 million yuan in damages. The case settled with a 500,000 yuan payment and forfeited salary. Tian describes the act as "fighting fire with fire": shutting down a colleague's process to free compute. The funding round shows the market has taken a side.

Underneath it sits a strategic bet: a 10x cut in video generation cost would change every robotics lab's data problem overnight. Synthetic training data at scale is the thing that could actually justify the brain valuations.

Common Pitfalls

• Assume your VLA understands paraphrases. VLM language robustness disappears in VLA policies. Test with adversarial phrasings before deployment, and consider a rule-based instruction rewrite at inference. The 16-27% relative gain on frozen policies is the cheapest improvement in this cluster.
• Quote post-trained benchmark numbers as zero-shot capability. VPP2's LIBERO-Pro, LIBERO-OOD, and RoboDojo top results come after benchmark-specific post-training. The zero-shot result is the ALOHA one. Match the evaluation to the claim.
• Read video prediction quality as action quality. A model that predicts pixels beautifully can execute horribly. VPP2 needed a separate mixture-of-transformers module to learn inverse dynamics. Without the action module, foresight never becomes motion.
• Fold every objective into one RL reward. End-to-end policies that couple gait, tracking, and energy in one reward are brittle and hypersensitive to reward weights. The HRL work decomposed the problem into a low-frequency gait command and a high-frequency tracking policy, then got zero-shot transfer and lower cost of transport. Decompose first.
• Price robots like software. The 74x multiple assumes Android-like distribution and software margins, but robot order volumes are still in the thousands. If you model a brain company's revenue, model the unit economics of the bodies it depends on.

One Thing to Remember

Across these five sources, one pattern holds: in embodied AI, the generalization story is the whole story. The rephrasing paper raised frozen-policy success by up to 27% with no new data. VPP2 closed the video-to-action gap with an inverse dynamics module. The locomotion team got zero-shot transfer by decomposing control. And the market is paying 74x revenue for the promise that any of it generalizes to the real world. Build the evaluation that hunts for the failure, the 2% phrasing, the 0.5 dB gain, the task your policy wasn't finetuned for. That evaluation is the actual product.

The Bottom Line

If you're deploying a VLA in production, add an instruction-rewrite layer at inference. Ten to twenty distilled rules, one rewrite per command, no retraining, and you get 16-27% relative gains on out-of-distribution tasks.

If you're building control for a legged platform, decompose the problem. A low-frequency gait policy that minimizes cost of transport, paired with a high-frequency tracking policy, beat end-to-end RL on energy and stability, and transferred zero-shot to a physical AlienGo.

If you're valuing a robot brain company, assume the Android multiple only pays off if robot volumes reach phone scale. Watch whether Tesla and BYD keep intelligence in-house. If they do, the independent brain ceiling is far below today's marks, and the next funding round will demand deployment metrics, not demos.