REVIEW 4 major objections 5 minor 3 references
Are Transformers Truly Foundational for Robotics?
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read Transformers give robots a facsimile of autonomy, not the real thing.
desk verdict A perspective piece with a genuinely useful insect-brain framing, but the central claim that transformers are structurally incapable of reasoning overreaches and needs major tempering. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the transformer block: a repeated unit of a self-attention mechanism followed by a multilayer perceptron, with feedforward control flow, that learns statistical associations from a very large training set. The paper counterposes it with the modular insect brain, whose specialised circuits (optic lobes, antennal lobes, mushroom bodies, and the central complex) exploit the regularities of the world to perform onboard navigation and decision-making with fewer than a million neurons. The contrast carries the argument: transformer generality comes from an undifferentiated architecture plus scale, whereas biological autonomy comes from structure-function specialisation combined with an elementary world model.
What would settle it
A single controlled benchmark where a transformer-based robot consistently reasons causally about a novel situation, gives a correct explanation of its action independent of training-set similarity, and maintains performance under adversarial sensor input would falsify the claim that transformers are structurally incapable of true autonomy.
Extended reading notes
Core claim
The paper's central claim is that transformers provide a facsimile of intelligent autonomy rather than true autonomy. A transformer robot resolves an input to an action by selecting the output that carries the highest predictive weight after training; there is no reasoning from a model of how the world works and no way to justify a choice other than statistical association to the training data. Because the architecture is feedforward and correlational, the authors argue that hallucinations are inescapable and reliable metacognition is structurally impossible, so transformer-based systems cannot serve as the foundation for robust robot autonomy. As an alternative, the paper points to the modular, structure-function-specialised insect brain, which solves navigation, foraging, and colony coordination with tiny compute, and proposes importing those design principles into robot control stacks.
Load-bearing premise
The load-bearing premise is that transformer weaknesses (hallucination, poor reasoning, high cost) are fundamental and cannot be engineered away; if future transformer variants add world models, reasoning loops, or verification modules, the core critique loses its force.
Editorial extensions
If this is right
- If transformers cannot reason, then transformer-based robots will require human teleoperation or isolation, which undercuts the promised economic benefits of autonomy.
- Embodied hallucination is more dangerous than in language-only models, and adversarial attacks on guardrails for LLM-controlled robots have already been demonstrated.
- Deployment costs will remain high: even reduced-size local models need on the order of 20-100 GB of memory, and compute and data requirements are both scaling exponentially.
- The 'bitter lesson' assumption that hand-designed structure always fails should be reconsidered; biological modularity may be a source of inductive bias that delivers robustness at low cost.
Reading between the lines
- Beyond the paper: if the correlational-limitation thesis is correct, then bolting a reasoning loop or verification module onto a transformer does not repair the foundation; the causal, model-based core must be the primary architecture.
- Beyond the paper: the bee-brain comparison suggests a concrete research programme—reverse-engineering circuits such as the central complex's ring attractor into low-power navigation modules and benchmarking them as replacements for large perception stacks.
- Beyond the paper: a testable extension would be to compare transformer-based and insect-inspired policies on the same embodied tasks while measuring out-of-distribution reliability, explainability, and energy per decision.
- Beyond the paper: because the paper is an opinion-review, the strongest version of its claim would require a systematic benchmark showing that transformer failures under novel conditions persist across scales and architectures.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This perspective paper questions whether transformer architectures, particularly generative pre-trained transformers (GPTs), should be treated as foundational for autonomous robotics. The authors argue that transformer-based systems entail prohibitive training and inference costs, depend on exponentially growing data, suffer from hallucinations and reasoning failures that they characterize as inherent, and at best provide a 'facsimile' of intelligent autonomy. They contrast this with the compact, efficient, and robust autonomy achieved by insect brains, and they propose alternative or complementary directions: Natural Intelligence (inspired by insect neurobiology), objective AI with explicit world models, and neurosymbolic hybrids. The paper concludes that the time is ripe to revisit the 'bitter lesson' and look to biological structure-function specialization for robotics.
Significance. If its strong claims were established, this paper would provide a rationale for redirecting substantial research effort away from transformer-centric robot learning and toward modular, biologically inspired architectures. The paper is timely and useful as a counterweight to the prevailing 'foundation model' narrative, and it correctly highlights real concerns about compute, data, and reliability that are widely acknowledged in the field. Its potential impact, however, is undercut by overgeneralization of contested theoretical results and by the absence of an operational definition of the very 'autonomy' it seeks to assess. The insect-brain comparison is thought-provoking and grounds the discussion in concrete biological evidence, but it is not yet a demonstrated engineering alternative.
major comments (4)
- [Transformers Give a Facsimile of Intelligent Autonomy] The central claim that transformer-based robots exhibit 'no reasoning' and are 'structurally incapable of reliable metacognition' conflates the bare feedforward transformer block described in Figure 1 with deployed systems that augment transformers with chain-of-thought prompting, retrieval, verifiers, external planners, and closed-loop interaction with the environment. The paper itself cites chain-of-reasoning models (ref 25) and then asserts that their failures are inherent, without an argument that such auxiliary mechanisms cannot overcome the alleged structural limits. To make the 'facsimile' claim load-bearing, the authors need an operational definition of the system class under critique and a direct engagement with published transformer-based robot systems that perform closed-loop manipulation or navigation; the current text relies on a survey (ref 7) and does not confront any concrete counterexample.
- [Hallucinations for Transformers in Robotics May Become Acute] The assertion that 'the fundamentally correlational nature of transformers will render hallucinations inescapable' is load-bearing for the paper's conclusion, but the evidence does not support the strong modal claim. Ref 23 is a theoretical result about a specific statistical setting, not an impossibility theorem for all transformer-containing systems with verification loops, world models, or external grounding; refs 19-20 and 24-25 document failures on current benchmarks, not architectural inevitability. Since the paper elsewhere acknowledges active mitigation research (ref 20), 'inescapable' should either be weakened to 'currently prevalent and difficult to suppress' or the authors should state the scope conditions of ref 23 and show that they apply to embodied robot controllers.
- [Introduction] The paper asserts that 'the level of autonomy of a robotics system is well-defined, measurable, and economically meaningful' but never supplies that definition or metric. The central distinction between 'true autonomy' and a 'facsimile of autonomy' cannot be adjudicated without an operational criterion. The authors should either provide a concrete measure (for example, intervention rate, robustness to distribution shift, or closed-loop task success under novel conditions) or reframe the argument as a comparative feasibility claim that does not depend on an absolute, and undefined, notion of 'true autonomy.'
- [Alternatives and Complements to Transformers for Autonomy] The Natural Intelligence proposal is presented with confidence that 'we don't need to match the process by which bee intelligence evolved if we want to match the performance of that evolved intelligence' and that this 'can be done by studying just the end point -- the embodied bee brain.' This is a research hypothesis, not a demonstrated result, and the paper provides no systems-level evaluation comparing the proposed approach against transformer-based alternatives. As a perspective piece, such a conjecture is acceptable, but it should be explicitly flagged as a conjecture; the current wording implies a superiority that has not been established.
minor comments (5)
- [Conclusion, ref 60] Reference 60 is cited for the 'bitter lesson' but lists the Sutton and Barto reinforcement learning textbook; the canonical source is Sutton's 2019 essay 'The Bitter Lesson.' The text also mischaracterizes the lesson as being about the failure of inductive biases, whereas Sutton's argument is that general methods that leverage computation tend to win in the long run.
- [References] Reference 28 ('Bommasani, R. et al. (2022)') is incomplete and does not match any entry in the reference list; the only Bommasani entry is ref 5 from 2021. Reference 25 ('Li, B. et al. (2024)') lacks a title and venue. These need to be completed.
- [Progress in Applying Transformer Architectures to Autonomy] The claim that 'one state of the art [object detection] algorithm has on the order of 10m-80m network weights' is uncited; ref 15 is the Llama 3.1 release note and does not support an object-detection weight count.
- [Figure 1 legend] The legend contains a typo: 'optic lopes' should read 'optic lobes.'
- [Conclusion] The first sentence contains a typo: 'muli-modal' should read 'multimodal.'
Circularity Check
No circular derivation: the critique is externally grounded and the authors' self-citations are not load-bearing.
full rationale
The paper makes no quantitative predictions and fits no parameters; it is a perspective argument. Its central claim—that transformers yield a "facsimile of autonomy" because they are feedforward, correlational, and lack explicit reasoning or self-justification—is supported by external evidence: Xu et al. (ref 23) for the inescapability of hallucinations, Mirzadeh et al. (ref 24) and Li et al. (ref 25) for reasoning failures, Bender et al. (ref 30) for the stochastic-parrot characterization, and Lake et al. (ref 31) for model-based reasoning. None of these is the authors' own result, and none is fitted or defined in terms of the conclusion. The authors' prior publications appear only in descriptions of insect neurobiology (e.g., Cope et al. 2017, Plath & Barron 2015) and in pointing to the insect-inspired AI research program (de Croon et al. 2022); these are not premises from which the transformer critique is derived. The Opteran affiliation is disclosed and raises a conflict-of-interest question, but it is not a circular argument. Whether the "structurally incapable" claim overreaches deployed transformer systems is a correctness concern, not a circularity concern.
Assumptions & free parameters
assumptions (4)
- domain assumption The fundamentally correlational nature of transformers makes hallucinations inescapable.
- domain assumption Feedforward transformer-based applications are structurally incapable of reliable metacognition.
- domain assumption Evolutionary pre-training has solved autonomy, so studying the endpoint (bee brain) gives a viable blueprint.
- domain assumption Transformers are 'unreasoned products of the training data' and lack world models.
Cite this review
Pith. "Pith review of Are Transformers Truly Foundational for Robotics?." pith.science (2026). https://pith.science/paper/JG73BI46
@misc{pith2026241116917,
author = {Pith},
title = {Pith review of: Are Transformers Truly Foundational for Robotics?},
year = {2026},
howpublished = {\url{https://pith.science/paper/JG73BI46}},
note = {Machine review of arXiv:2411.16917}
}
read the original abstract
Generative Pre-Trained Transformers (GPTs) are hyped to revolutionize robotics. Here we question their utility. GPTs for autonomous robotics demand enormous and costly compute, excessive training times and (often) offboard wireless control. We contrast GPT state of the art with how tiny insect brains have achieved robust autonomy with none of these constraints. We highlight lessons that can be learned from biology to enhance the utility of GPTs in robotics.
Reference graph
Works this paper leans on
-
[1]
Are Transformers Truly Foundational for Robotics? James A. R. Marshall1,2,* and Andrew B. Barron3 1Centre for Machine Intelligence and Department of Computer Science, University of Sheffield, S10 2TN, UNITED KINGDOM 2Opteran Technologies Ltd., The Innovation Centre, 217 Portobello, S1 4DP, UNITED KINGDOM 3School of Natural Sciences, Macquarie University, ...
work page 2019
-
[2]
Transformer-based approaches to autonomy rely on internet-scale datasets as input, and are trained to process input from a suite of high resolution sensors, such as 4k cameras and LiDar, in order to provide a limited behavioural repertoire, in comparison to biological autonomous agents. In contrast, 600m years of evolution on a planetary scale, with compl...
arXiv 2023
-
[56]
However, this results in an autonomy solution very different to the way the only truly autonomous artefact known to humanity, the biological brain, functions. Here we have highlighted this, and conclude by arguing that the tremendous recent advances in data on, and understanding of, a variety of brains, means the time is ripe to revisit the ‘bitter lesson...
work page 2020
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.