Embodied AI in Machine Learning -- is it Really Embodied?

Matej Hoffmann; Shubhan Parag Patni

REVIEW 3 major objections 5 minor 2 references

Embodied AI in Machine Learning -- is it Really Embodied?

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Today's AI robots inherit GOFAI's core failings, a new chapter argues.

desk verdict A useful conceptual critique that applies old embodiment principles to recent foundation-model robotics; the 'roadblock' claims are best read as clearly-flagged conjectures, not established results. read the letter →

arxiv 2505.10705 v1 pith:3KFMLKIO submitted 2025-05-15 cs.AI cs.NEcs.RO

Matej Hoffmann , Shubhan Parag Patni This is my paper · ORCID

classification cs.AIcs.NEcs.RO

keywords embodiedAIweaklyGOFAIlargelanguagemodelsrobotlearningcross-embodimentsensorimotorcoordinationfoundation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This chapter seeks to establish a negative thesis about the current 'Embodied AI' wave in machine learning: robots controlled by large language and vision-language models are only weakly embodied, and they inherit the same fundamental problems that sank Good Old-Fashioned AI (GOFAI). The authors evaluate representative architectures—LLM planners, VLM state estimators, code-generating action modules, and their fusions such as RT-1, RT-2, and PaLM-E—against an embodiment checklist derived from the behavior-based robotics tradition. Their conclusion is that the body is treated as an abstracted periphery (a low-dimensional end-effector action space), so morphology, sensorimotor coordination, active perception, and multiple interaction time scales are not exploited. The stakes are concrete: if the paper is right, scaling up datasets and model size will not by itself produce robust real-world robotics, and the field must shift toward active learning, embodiment-aware representations, and co-design of bodies and controllers.

What carries the argument

The central device of the paper is the embodiment checklist, five principles drawn from Pfeifer and Scheier (2001) and Pfeifer and Bongard (2006): (1) body morphology facilitates control, (2) sensor morphology facilitates perception, (3) sensorimotor coordination and active perception, (4) parallel loosely coupled processes, and (5) the principle of ecological balance between morphology, materials, control, and environment. The authors use this checklist as a measuring stick, applying it to representative WEAI systems such as SayCan, Inner Monologue, Code as Policies, PaLM-E, and RT-X. The comparison is anchored by a figure contrasting the implications of embodiment—where mechanical feedback and body dynamics participate in generating information structure—with the WEAI architecture, where the body is reduced to an arrow closing the loop between sensors and actuators. The checklist does the argument's work: it converts an intuitive complaint about 'shallow embodiment' into a point-by-point diagnostic.

What would settle it

Find a robot system trained entirely on offline teleoperation data (no active interaction during learning) that robustly solves a broad suite of novel real-world manipulation and locomotion tasks requiring sensorimotor adaptation and physical interaction beyond its training distribution; or show that a cross-embodiment model trained on dozens of diverse robots reliably transfers zero-shot to a completely new morphology with different limb structure and sensor layout. Either demonstration would directly contradict the paper's claim that passive, embodiment-abstracting learning is fundamentally limited.

Watch

Extended reading notes

Core claim

On the authors' own terms, the discovery is that the contemporary 'Embodied AI' research program, for all its surface novelty, is structurally a modern, subsymbolic sense-think-act architecture. Reasoning is performed in a large model; the world is represented in language or latent embeddings; state re-estimation and replanning are needed to keep that representation in sync with the world, the frame problem is managed rather than dissolved; and grounding is inherited indirectly from human-centered internet text and images. Measured against the embodiment principles of Pfeifer, Brooks, and the behavior-based tradition—behavior arising from closed-loop interaction, morphology facilitating control, sensor morphology shaping perception, sensorimotor coordination, parallel loosely coupled processes, and ecological balance—the authors conclude that WEAI (weakly embodied AI) scores poorly on every point. The claim is not that these systems fail at present, but that their embodiment is shallow, their data collection is passive, and their cross-embodiment efforts succeed only because the embodiments have been abstracted to Cartesian actions, which puts a principled ceiling on what scaling alone can achieve.

Load-bearing premise

The whole critique hangs on accepting that the embodiment principles from the behavior-based and cognitive-science tradition—morphology must facilitate control and perception, and active sensorimotor interaction is constitutive of intelligence—are the right yardstick for judging robotic intelligence; if one believes a controller can in principle be intelligent while its body is only a generic interface, the paper's charge of 'weak embodiment' loses much of its force.

Editorial extensions

If this is right

If the paper's thesis is correct, then the current practice of collecting large teleoperation datasets and training vision-language-action models on them will yield controllers that fail to exploit even available morphological and interactive resources, because the data collection itself is passive.
Cross-embodiment learning would face a principled trade-off: the more diverse the embodiments and their sensory and action spaces represented in one model, the less each individual robot can profit from the shared 'brain', so positive transfer will shrink as the embodiment gap grows.
To make use of embodiment, models must either include multiple interaction loops at different time scales and more mechanical and sensory detail, or bypass explicit modeling and let controllers learn active closed-loop interaction directly, as in deep reinforcement learning.
The implication for evaluation is that systems should be judged not only on task success rates but on whether they genuinely exploit sensorimotor coordination, morphology, or active perception—factors that current benchmarks do not isolate.
The direction of travel implied by the paper is a return to single-embodiment active learning, plus the use of large models only for the parts where they are indubitably useful (flexible planning and passive visual perception), with a robot-control API handling real-time interaction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable prediction that follows from the paper's logic: at matched data volume, an egocentric, actively replanning policy that can physically sample the environment will outperform a passively trained offline policy on tasks requiring visuomotor adaptation, paralleling the Held and Hein kitten experiment.
The argument implies that benchmark progress in 'Embodied AI' may be measuring something narrower than embodied intelligence—specifically, the robustness of language and image priors plus the capability of low-level motion controllers—so a hidden confound exists in current leaderboards.
If cross-embodiment transfer becomes a core benchmark, the field may be forced to choose between evaluating generalization across genuinely different bodies (where the paper predicts limited transfer) and standardizing a single body (where transfer is trivial), and the paper's position suggests the latter is the only reliable road to immediate gains.
One could extend the authors' ecological-balance principle into a quantitative design criterion: the information-theoretic complexity of the controller should be matched to the complexity of the body and environment; current 'gigantic brain, tiny action space' configurations are, by this measure, ecologically imbalanced by construction.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The one thing to know: this is a book chapter, not a research paper, and it reads like one. No new experiments, no new derivations. What it does well is give a clean vocabulary for a real phenomenon. The label 'weakly embodied AI' is genuinely useful, and the checklist derived from Pfeifer and Bongard is a practical tool for anyone trying to evaluate claims about embodied AI. I plan to borrow it. The chapter is also fair-minded: it walks through RT-2, Open X-Embodiment, and related work without caricature, and it correctly notes that cross-embodiment generalization to new robots had not yet been studied in the main reference. The Held and Hein kitten example is well placed and not oversold. The citation pattern looks honest, with self-citations only on secondary points. Where it gets soft is the step from 'weakly embodied' to 'fundamental roadblocks.' The claim that there is a fundamental trade-off between embodiment diversity and shared-model benefit is argued qualitatively; the authors themselves admit the decisive experiment is missing. So the roadblocks are plausible conjectures, not established limits. The stress-test note is right about that, but it is also a bit harsh: the chapter flags the missing evidence in the text, so it is not sweeping the problem under the rug. A more serious limitation is that the whole critique rests on a normative premise—that Pfeifer and Brooks are the right standard for judging intelligence and robotics. Readers who do not share that premise will not be convinced, and the chapter does not try to defend it. That is fine for a chapter aimed at the faithful, but it limits the reach. The generalizations are also based on a small set of systems, though the selection is at least representative. In short: this is a solid, clearly written conceptual critique that will be useful to people in ML robotics and embodied cognition. It deserves serious peer review, and I would cite it as a counterweight to the hype. The main recommendation is that the chapter should be read as opening a debate, not closing one.

Referee Report

3 major / 5 minor

Summary. This book chapter argues that the current wave of large-model-driven robotics, which the authors call Weakly Embodied AI (WEAI), is only weakly embodied and inherits some of the problems of GOFAI. The authors define an embodiment checklist based on Pfeifer, Brooks, and related work, apply it to representative systems (e.g., GATO, SayCan, PaLM-E, RT-1/2), then review cross-embodiment datasets and models (Open X-Embodiment, J. Yang et al. 2024, "All Robots in One"). They conclude that there are fundamental roadblocks to scaling current foundation-model approaches to robotics, in particular a trade-off between embodiment diversity and transferable benefit, and the insufficiency of passive offline data. They propose directions such as egocentric perception, robot-aware affordances, and combining foundation models with low-level embodied controllers.

Significance. If the central thesis is accepted, the chapter is a valuable and timely counterweight to the prevailing assumption that scaling data and compute will solve robot learning. Its strengths are: (i) a clear and well-grounded conceptual framework that connects the behavior-based/embodied tradition with current ML practice; (ii) an honest presentation of the current evidence, including the admission that RT-X generalization to new robots has not yet been studied; and (iii) concrete, testable recommendations (e.g., egocentric data collection, robot-aware affordances). However, the paper's strongest conclusion—that WEAI will hit "fundamental roadblocks"—depends on empirical conjectures about cross-embodiment transfer and passive data that are not yet demonstrated. The chapter is therefore best read as a persuasive position piece; the load-bearing empirical claims need either direct evidence or explicit reframing as hypotheses.

major comments (3)

[New foundation models for robotics?] The paragraph beginning "There is thus a fundamental trade-off" asserts a "principled limitation of foundation models for robotics", namely that the more diverse the robot embodiments, the less individual robots can profit from a shared model. This is presented as a settled result, but the paper itself notes in the previous section that for Open X-Embodiment/RT-X "generalization to new robots was not yet studied". No quantitative or formal support is given for the trade-off. Because this claim is load-bearing for the paper's central conclusion that WEAI will face fundamental roadblocks, I ask the authors to either (a) soften the claim to an explicit empirical hypothesis, or (b) provide the missing evidence, e.g., a summary of the zero-shot generalization results from J. Yang et al. (2024) or a systematic analysis of transfer versus embodiment distance.
[Active embodied interaction versus offline learning] The inference that robots trained on passive teleoperated datasets are "bound to be inefficient" and that active learning is "necessary" relies on an analogy to Held and Hein (1963). The authors themselves note a crucial disanalogy: robot datasets typically contain both sensory and motor signals, unlike the passive kitten, which lacked motor commands. The Held-Hein experiment is also about developmental plasticity in kittens, not about statistical learning from supervised offline data. As written, the jump from this analogy to a "fundamental limitation" in the section "New foundation models for robotics?" is an empirical conjecture, not a demonstrated result. Please either temper the claim or support it with direct evidence from robot-learning studies.
[Cross-embodiment Embodied AI – an oxymoron?] The discussion of J. Yang et al. (2024) states that zero-shot generalization to a new embodiment "was evaluated" but does not report any outcome. This is the single most relevant experiment for the paper's roadblock thesis: it directly tests whether transfer fails as embodiment distance grows. The chapter should at least summarize the qualitative finding (e.g., success on similar embodiments, failure on distant ones) or quote the relevant result, so that readers can judge whether the "fundamental trade-off" is supported.

minor comments (5)

[GOFAI-powered robots never really worked] The citation "Haugeland(1989)" should have a space before the year and the closing parenthesis.
[Representative works] The sentence "Examples of this approach are RT-2 (Brohan et al. 2023) or VC-1 (Yokoyama et al. 2023)" misattributes the cited Yokoyama et al. (2023) paper, whose title is "ASC: Adaptive Skill Coordination for Robotic Mobile Manipulation"; please correct the model name or the reference.
[Where WEAI inherits GOFAI problems] The claim that closing the loop at "3-10 Hz ... is not fast enough for real tasks in robotics" needs a citation or a more careful qualification, since many manipulation and navigation tasks are executed at lower rates and performance depends on the control hierarchy.
[How embodied is Weakly Embodied AI?] The statement that WEAI architectures have "only one process and one time scale" may be too strong; for example, some systems combine a slow planner with a fast reactive policy. Consider softening this claim.
[Figure 3 caption] The phrase "A cartoon-like representation dataset generation in WEAI" appears to be missing a preposition (e.g., "of"); please revise the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central weak-embodiment claim is an evaluative application of an external Pfeifer/Bongard checklist; the self-citations are secondary and not load-bearing.

full rationale

The chapter's central claim—that current AI-powered robots are only weakly embodied—is an evaluative judgment made against an external checklist drawn from Pfeifer and Scheier (2001) and Pfeifer and Bongard (2006), not from the WEAI models themselves. The checklist items (body morphology facilitates control, sensor morphology facilitates perception, sensorimotor coordination and active perception, parallel loosely coupled processes, ecological balance) are independent normative criteria; the paper then checks representative WEAI architectures (GATO, SayCan, RT-1/RT-2, RT-X, J. Yang et al.) against those criteria. No parameter is fitted, no prediction is made from a subset of data, and no known result is renamed as a new principle. The self-citations (Hoffmann and Pfeifer 2012; Müller and Hoffmann 2017; Pliska et al. 2024) support only secondary points about mechanical feedback, the terminology of morphological computation, and gripper-dependent tactile data; none of these citations carries the load of the central weak-embodiment claim. The 'fundamental roadblocks' arguments (the embodiment-diversity/transfer trade-off and the insufficiency of passive offline data) are empirical conjectures rather than circular reductions; the paper itself flags the missing evidence, noting that for Open X-Embodiment 'generalization to new robots was not yet studied' and reporting that J. Yang et al. (2024) evaluated zero-shot generalization without giving quantitative results. That is a support gap for the correctness of the roadblock claim, not a circularity of derivation. The only possible circularity concern—that 'weakly embodied' is true by definition of the chosen checklist—does not vitiate the argument, because the checklist is not derived from WEAI and the conclusion is a substantive application of an external standard. No circular step is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper's argument relies on the embodiment framework from Pfeifer, Brooks, and ecological psychology as a normative standard. These are domain assumptions the reader must accept for the critique to land; there are no fitted parameters or invented entities.

assumptions (4)

domain assumption Intelligent behavior arises from closed-loop interaction of controller, body, and environment, not from the controller alone.
The paper's checklist in 'Implications of embodiment' uses this as the standard to judge WEAI; it is taken from Pfeifer and Scheier without independent verification.
domain assumption Body morphology and materials can substantially simplify control and perception.
Assumed in the checklist principles 1 and 2; supported by passive dynamic walking and compound-eye examples, but treated as a general law.
domain assumption Active interaction is necessary for proper sensorimotor development and learning (Held and Hein 1963 results generalize).
Used in 'Active embodied interaction versus offline learning' to argue offline teleoperation datasets are fundamentally limited.
domain assumption The internet text and image corpora are human-centric and thus provide only shallow, indirect grounding for robots.
Underpins the symbol grounding critique; plausible but asserted rather than demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Embodied AI in Machine Learning -- is it Really Embodied?." pith.science (2026). https://pith.science/paper/3KFMLKIO

@misc{pith2026250510705,
  author       = {Pith},
  title        = {Pith review of: Embodied AI in Machine Learning -- is it Really Embodied?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3KFMLKIO}},
  note         = {Machine review of arXiv:2505.10705}
}

read the original abstract

Embodied Artificial Intelligence (Embodied AI) is gaining momentum in the machine learning communities with the goal of leveraging current progress in AI (deep learning, transformers, large language and visual-language models) to empower robots. In this chapter we put this work in the context of "Good Old-Fashioned Artificial Intelligence" (GOFAI) (Haugeland, 1989) and the behavior-based or embodied alternatives (R. A. Brooks 1991; Pfeifer and Scheier 2001). We claim that the AI-powered robots are only weakly embodied and inherit some of the problems of GOFAI. Moreover, we review and critically discuss the possibility of cross-embodiment learning (Padalkar et al. 2024). We identify fundamental roadblocks and propose directions on how to make progress.

Figures

Figures reproduced from arXiv: 2505.10705 by the authors.

**Figure 2.** Evolution of approaches in Embodied AI. Redrawn and adapted from (Vanhoucke [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Active embodied interaction versus being driven around. (A) Setup from (Held and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: To exploit [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 linked inside Pith

[36]

All Robots in One: A New Standard and Unified Dataset for Versatile, General-Purpose Embodied Agents

PMLR. Walter, G. W. 1953. The Living Brain. Edited by G. W. Walter. Norton & Co., New York. Wang, Zhiqiang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, et al. 2024. “All Robots in One: A New Standard and Unified Dataset for Versatile, General-Purpose Embodied Agents.” arXiv. http://arxiv.org/abs/2408.10899. Yang, Fengyu, Chao Feng, ...

arXiv 1953
[2024]

Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI

“Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI.” arXiv. http://arxiv.org/abs/2407.06886. Lungarella, Max, and Olaf Sporns. 2006. “Mapping Information Flow in Sensorimotor Networks.” PLOS Computational Biology 2 (10): e144. https://doi.org/10.1371/journal.pcbi.0020144. McGeer, Tad. 1990. “Passive Dynamic Walking.” The Inte...

arXiv 2006

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [36]

All Robots in One: A New Standard and Unified Dataset for Versatile, General-Purpose Embodied Agents

PMLR. Walter, G. W. 1953. The Living Brain. Edited by G. W. Walter. Norton & Co., New York. Wang, Zhiqiang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, et al. 2024. “All Robots in One: A New Standard and Unified Dataset for Versatile, General-Purpose Embodied Agents.” arXiv. http://arxiv.org/abs/2408.10899. Yang, Fengyu, Chao Feng, ...

arXiv 1953

[2] [2024]

Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI

“Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI.” arXiv. http://arxiv.org/abs/2407.06886. Lungarella, Max, and Olaf Sporns. 2006. “Mapping Information Flow in Sensorimotor Networks.” PLOS Computational Biology 2 (10): e144. https://doi.org/10.1371/journal.pcbi.0020144. McGeer, Tad. 1990. “Passive Dynamic Walking.” The Inte...

arXiv 2006