Language models learn to write essays without anyone teaching them sentence by sentence. The entire internet serves as their free textbook—every piece of text is already written. But robots learning to install batteries have no ready-made textbook: each grasp, each application of force, must be physically performed and recorded by a human in the real world to generate a single training data point. Behind this asymmetry lies an old problem that has plagued AI for decades. As early as 1988, roboticist Hans Moravec pointed out in *Mind Children*: what is hard for machines is precisely what is easy for humans—perception and movement; and what is easy for machines is precisely what is hard for humans—reasoning. This observation later became known as Moravec’s paradox. In an era where large models have turned "reasoning" into a commodity, the other half of the paradox has instead grown increasingly expensive.
"Embodied AI" refers to intelligence that must acquire information and act upon the physical world through a physical body—it is both the next step beyond traditional robots (which rely on hand-coded programming) and the mirror image of software agents (which live entirely in digital environments). From 2023 to 2024, a wave of papers emerged densely in robotics, pursuing diverse technical paths—but read side-by-side, they converge on the same conclusion: what stalls embodied intelligence is not algorithms, but data—more precisely, the lack of low-cost, scalable channels for acquiring physical interaction data. Language has the internet; physical interaction does not.
## Language Models Have the Internet—What Do Robots Have?
The Open X-Embodiment paper bluntly describes the traditional approach to robot learning at its outset: train a separate model for each application, each robot, and even each environment. In other words, data cannot be reused—experience accumulated by Robot A is worthless to Robot B.
Why don’t language models need to follow this path? Because their data acquisition cost approaches zero: text is produced by humans for human purposes, and AI merely picks it up opportunistically. Robot action data is precisely the opposite: it does not exist naturally and must be deliberately produced—either by humans guiding robots’ hands in demonstrations, or by letting robots repeatedly try and fail in the real world. Production incurs cost—and that cost cannot be amortized.
This is the true meaning of Moravec’s paradox in the era of large models: the bottleneck is not the "body" itself, but the absence of scalable channels for acquiring bodily data. Around this gap, four data-oriented pathways have successively emerged: making data cheaper to produce, stitching data together, transferring knowledge from the internet into action, and reusing old data for new tasks. Let’s examine them one by one.
## Pathway One: Making Data-Production Hardware Cheaper
In April 2023, a paper titled ALOHA (arXiv:2304.13705) posed a direct question: "Can learning enable low-cost and imprecise hardware to perform these fine manipulation tasks?"
Fine manipulation—such as threading zip ties or installing batteries—traditionally requires high-end robots, precision sensors, or meticulous calibration, all of which are expensive and difficult to set up. ALOHA takes the opposite approach: using low-cost, imprecise hardware paired with a custom teleoperation interface for direct human demonstration, followed by end-to-end imitation learning.
Imitation learning is tricky: small policy errors compound over time, and human demonstrations themselves are non-stationary. ALOHA’s algorithm, ACT (Action Chunking with Transformers), avoids predicting the next action step-by-step. Instead, it learns a generative model over action sequences and outputs a chunk of actions at once. Result: the robot learned six challenging real-world tasks—including opening a translucent spice jar and installing batteries—with success rates of 80–90%, using only 10 minutes of demonstration data.
The essence of this pathway is not training stronger models, but driving down the unit cost of producing a single demonstration to an acceptable level; when data becomes cheap, algorithmic shortcomings can be compensated for with volume.
## Pathway Two: Aggregating All Robots’ Data Together
If data from a single robot is insufficient, what about aggregating data from all robots? Open X-Embodiment (arXiv:2310.08864, first released in October 2023) asks precisely this question. Its abstract poses it directly: "Can such a consolidation happen in robotics?"—NLP and computer vision have already traversed the path of "large-scale diverse data → pre-trained general-purpose backbone → support for downstream applications." Can robotics replicate it?
Their approach begins with manual labor: through collaboration among 21 institutions, they aggregated data from 22 different robot embodiments, standardized it into a unified data format, covering 527 skills and 160,266 tasks. On this foundation, they trained a high-capacity model, RT-X, and observed positive transfer: robots improved their performance by leveraging experience learned on other platforms.
This marks the first partial replication in robotics of NLP’s "pre-training consolidation" paradigm—a cautious affirmative answer to the question posed in the abstract. But the prerequisite is explicit: data formats must first be standardized—otherwise, aggregation is impossible. And the word "partial" is not mere modesty: the cross-embodiment gap has not disappeared.
## Pathway Three: Transferring Internet Knowledge into Action
Physical data is expensive; internet data is free—so can we transfer knowledge already acquired by models from the internet into robot actions? In July 2023, RT-2 (arXiv:2307.15818) proposed an elegantly minimal recipe: represent actions directly as text tokens, embed them in the same format as natural language, and then co-fine-tune a state-of-the-art vision-language model jointly on robot trajectory data and internet-scale vision-language tasks (e.g., visual question answering).
They named such models VLAs (vision-language-action models). Across 6,000 evaluation trials, RT-2 demonstrated emergent capabilities attributable to internet-scale training: significantly improved generalization to novel objects, and the ability to understand instructions never seen in its robot training data—for example, placing an object onto a specific number or icon.
The essence of this pathway is "outsourcing": semantic commonsense (e.g., what a number is, what an icon is, what "place on top" means) is outsourced to internet corpora, while expensive physical demonstration data is reserved solely for mapping actions. The more commonsense that can be outsourced, the less needs to be taught via physical demonstration.
## Pathway Four: Leveraging Old Data for New Tasks
In January 2024, Mobile ALOHA (arXiv:2401.02117) extended ALOHA with a mobile base and full-body teleoperation interface, enabling collection of mobile manipulation data requiring bimanual plus whole-body control. Yet its contribution to the data pathway is an even more cost-effective insight: when training for new tasks, co-train with existing static ALOHA data—each new task requires only 50 demonstrations, and co-training boosts success rates by up to 90%.
Using this recipe, robots can successfully pan-fry a shrimp and serve it on a table, open a double-door cabinet and stow a heavy pot inside, call and enter an elevator, and briefly rinse a used frying pan under a faucet. Note the figure "50 demonstrations": the marginal cost of new tasks is further reduced—old data becomes a shared foundational repository, and new tasks require only a thin layer of incremental demonstrations on top.
The essence of this pathway is reuse: data is no longer a single-use consumable; every previously collected demonstration continues working for new tasks.
## Convergence—and My Three Judgments
In October 2024, the π0 paper (arXiv:2410.24164, last revised in January 2026) integrated all four pathways. It builds a flow-matching architecture atop a pre-trained vision-language model, inheriting internet-scale semantic knowledge—this is Pathway Three; its training data comes from multiple dexterous robot platforms—single-arm, bimanual, and mobile manipulation—large in scale and diverse in type—this merges Pathways One and Two; and it supports acquiring new skills via fine-tuning—echoing Pathway Four. Tasks like folding clothes, clearing tabletops, and assembling cardboard boxes entered the evaluation list. Even so, the abstract acknowledges: "major obstacles in terms of data, generalization, and robustness remain before robot learning achieves the generality required for real-world systems."
The following three judgments are my synthesis drawn from reading all five papers side-by-side—not conclusions stated in any single original paper.
First, the current competition in embodied intelligence is a contest of data strategy, not model architecture (my judgment). Whether ACT’s action chunking or π0’s flow matching, architectural innovations are fundamentally tools to "make limited data more usable"; the real decisive factor is who can produce physical interaction data more cheaply, aggregate it more centrally, and reuse it more repeatedly. All four pathways—production, aggregation, transfer, and reuse—are data economics, not algorithmics.
Second, robotics has only "partially" replicated the pre-training paradigm (standard interpretation). OXE’s positive transfer proves that cross-embodiment data aggregation is useful—but aggregation presupposes format standardization, and the returns are bounded: π0 still requires platform-specific data tuning. The cross-embodiment gap has narrowed, but not vanished; "one model fits all robots" remains a hypothesis, not an engineering reality.
Third, the inversion of Moravec’s paradox determines AI’s commercialization sequence (industry observation). Software agents interact digitally—their data accumulates near-zero-cost automatically during use; embodied agents incur production cost for every action datum. So the dividing line is not "intelligence level," but "data acquisition cost per unit of intelligence"—software agents lead in commercialization, irrespective of how "smart" they are.
## What to Take Away
Next time you see news claiming "robots achieve human-like manipulation," assess its substance with three questions: How many demonstration data points underlie the reported success rate—10 minutes, 50 demonstrations, or large-scale collection? Can this data be reused across tasks and across robot embodiments—or is it one-time-only? Does the model’s semantic commonsense come from internet transfer, or from pure physical trial-and-error?
These three questions respectively probe data production cost, aggregation and reuse, and knowledge transfer. For general readers, there is one further takeaway worth remembering: the real barrier to embodied intelligence is not "how smart the robot is," but the economics of data. Whoever first makes physical interaction data as cheap and tradable as text will seize robotics’ "pre-training moment."
Among the four data pathways, which do you think will succeed first? Feel free to share your thoughts in the comments.
Source: Open X-Embodiment paper arXiv page https://arxiv.org/abs/2310.08864