{"id":"7342a26a-849d-49d7-9046-49c2a1ddd3ce","arxiv_id":"2611.02510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A brain-inspired design proposal: compliant control, a fast vision stream, a self-calibrating predictor, and primitives consolidated from code into weights, offered as a hypothesis for closing the human-robot skill gap.","lead":"This position paper argues that robot policies fail at everyday manipulation because of architecture, not data scale, and proposes a brain-inspired blueprint: compliant action, a fast vision channel for control, a predictor that forecasts outcomes, and primitives that are practiced and consolidated into weights. It is worth reading because it bundles several active research threads into one testable design for a robot that keeps improving after deployment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Deployment-time self-improvement rests on an autonomous judge that the paper itself says current models lack; the practice/consolidation agents assume this judgment instead of providing it.","rationale":"The reader's verdict is CONDITIONAL, and my read does not move it. The paper is a self-described position paper (b575-b576), labels its neuroscience readings, and offers each module as a hypothesis; I found no fabrication, circular derivation, or unsupported quoted theorem. The load-bearing weak point is the third of the three abilities the abstract and Section 1 center on: 'keep improving.' Unlike compliance, the fast vision stream, and the predictor, whose engineering counterparts are concrete (admittance/impedance, a second small encoder, a world model), the self-improvement loop is specified only by naming agents—practice, judge, consolidation—and assigning them duties. The hardest duty, autonomous diagnosis of which primitive failed and which correction moves closer to success, is precisely what the paper says current models 'rarely manage' (Section 6.3.1) and what Section 6.1.5 says still falls to people (b402). So the central claim's most ambitious clause rests on a capability that is asserted, not designed. This supports CONDITIONAL rather than ACCEPT. It does not justify REJECT: the proposals are concrete enough to be built and measured, and the paper's own Figure 27 specifies a testable instance of the judge's task. My concrete_test uses that instance. I agree with the reader's weakest_assumption in part: the same frontier-model capability sits underneath my concern, but my emphasis is narrower—not whether the analogy transfers, but whether the autonomous judge exists at all. Hence agreement_with_reader: partial.","tokens_in":56568,"tokens_out":7522,"duration_ms":52026,"concrete_test":"Re-run the paper's own worked example of Figure 27 as a pass/fail benchmark: give a frontier multimodal model (e.g., GPT-6 Astra) the seven-episode boundary-test transcripts for the lock task, including forecast-versus-outcome mismatches and model-call counts, and require it to (1) name orient.angle as the error-prone unit and (2) output the corrective halving interval that terminates at 37 degrees, without human labels. If the model cannot do both, the autonomous judge on which the 'keep improving after deployment' leg rests is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The third leg of the central claim (b2) is that the proposed robot 'keep[s] improving after deployment.' That leg depends on an autonomous evaluation loop: the practice agent selects the weakest unit by attributing failures across units, the judge scores boundary outcomes, and the consolidation agent verifies that a proposed update is a real gain and decides what to write to the weights (Sections 6.3.4, 6.3.6, 6.3.7). The paper itself identifies the core of this loop as open: Section 6.1.5 states that 'choosing what to attempt next, deciding when a weight update is warranted, and verifying that an update is a real gain still fall to people' (b402), and Section 6.3.1 acknowledges that current multimodal models 'rarely manage this diagnosis, least of all where forces interact out of the camera's view.' The proposals do not supply a mechanism for these judgments. Section 6.3.7 says correction 'rests on one judgment: which of two outcomes lies closer to the goal,' and the conclusion proposes to 'strengthen that judgment,' but no training objective, architecture, or data source for the judge is specified beyond typed experts (Section 4.3.2). A system with compliance, a fast control encoder, and a self-labeling predictor could be built exactly as proposed and still not improve after deployment if the judge cannot attribute blame and score progress. The paper offers this as a hypothesis to be tested (b576), which is honest, but the central claim's self-improvement leg therefore has no demonstrated mechanism; the practice and consolidation agents restate the requirement rather than satisfy it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper is a survey and position paper. It argues that the gap between current robot policies and human sensorimotor competence is one of design rather than data scale, and it proposes a next-generation robot architecture assembled from five borrowings from the nervous system: (i) commanding stiffness alongside position, realized beneath the policy by impedance or admittance control and modulated by touch (Sections 2 and 3); (ii) a dedicated fast streaming vision encoder for control running beside the recognition encoder (Section 4); (iii) a self-labeling, shallow-wide-sparse predictor beside the policy, modeled on the cerebellum, that recalibrates from its own error (Section 5); (iv) a single multimodal model that composes parameterized action primitives into action units verified at sensory boundaries (Sections 6.3.1 through 6.3.3); and (v) practice and consolidation agents that graduate proven primitives from code into weights, intended to produce deployment-time self-improvement (Sections 6.3.4 through 6.3.7). The stated thesis (b2) is that a robot so organized would generalize to unfamiliar tasks, react at camera rate, and keep improving after deployment. The authors are explicit that the neuroscience is partly contested, that the analogy has limits, and that each proposal is offered as a hypothesis to be tested (b575-b576).","tokens_in":56918,"tokens_out":17200,"duration_ms":105158,"significance":"If the design thesis holds, the paper makes a useful contribution to the active debate among VLA, world-model, and code-as-policy architectures: it offers a coherent alternative with clean module boundaries, and several modules are concrete enough to build or falsify today, including the force-field touch representation (Section 3.3.1), the streaming control encoder (Section 4.3.1), body-part embeddings (Section 5.3.4), and the code-to-weights consolidation path (Sections 6.3.2 and 6.3.4). The survey reporting is careful: contested neuroscience is flagged (the dorsal/ventral division in Sections 4.2.1 and 4.2.2; the forward-model reading of the cerebellum in Section 5.2), quantitative claims are cited to sources (for example, OmniTacTune results on contact-rich tasks in Section 2.3.2 and GPT-6 Astra zero-shot success in Section 6.1.4), and the paper separates its own proposals from established results. The explicit limitation statements (b575-b576) strengthen trust in the survey portions, and the module-level proposals are genuinely falsifiable predictions.","major_comments":[{"comment":"The abstract's third advertised capability, \"keep improving after deployment\", is not entailed by the proposed components. The practice agent (Section 6.3.6) and the consolidation agent (Section 6.3.4) both depend on a graded judgment, \"which of two outcomes lies closer to the goal\" (Section 6.3.7), but the paper supplies no training objective, architecture detail, or data source for that judge beyond the typed experts of Section 4.3.2, and the Conclusion only proposes to \"strengthen that judgment\". The paper itself states in Section 6.1.5 (b402) that \"deciding when a weight update is warranted, and verifying that an update is a real gain still fall to people\", and Section 6.3.1 concedes that current multimodal models \"rarely manage this diagnosis\". Because the judge is a prerequisite for both the weakest-unit selection and the update-acceptance decision, this leg does not follow from the compliant arm, fast encoder, and self-labeling predictor, and a system built exactly as proposed could fail it. The attribution half of the loop is reasonably specified by the three observable signs in Section 6.3.6, and the boundary tests of Section 6.3.1 provide binary pass/fail signals, but no graded goal-distance signal is identified. I recommend either specifying a concrete learning signal for the judge (for example, bootstrapping from boundary-test outcomes and forecast-versus-outcome discrepancies as described in Sections 3.3.2 and 5.3.1, or from the success-scored rollouts reviewed in Section 6.1.2) or explicitly demoting the self-improvement claim from the abstract's central triad to the paper's most speculative proposal.","section":"Abstract; §6.3.7","authors_acknowledge":"Automating the loop is the open part.The three routes are automated unevenly: agents already judge their own outcomes and write skills back, while choosing what to attempt next, deciding when a weight update is warranted, and verifying that an update is a real gain still fall to people.","authors_acknowledge_p":0.69},{"comment":"The central claim is stated as a joint consequence of the assembled design, but the paper never states which component is necessary for which of the three advertised capabilities. The camera-rate leg depends on the streaming encoder of Section 4.3.1 together with the reactive head of Section 6.3.3; the unfamiliar-task leg depends on the primitive library and code path of Section 6.3.1; and the self-improvement leg depends on the judge (major comment above). The paper also contains components it explicitly labels as open (\"The exact design ... is to be determined\", Section 5.3.3; \"drawn with wide imagination\", Section 6.3.5; \"No model today infers fast enough for this\", Section 5.3.2), so the reader cannot tell which parts of the tripartite claim are load-bearing. Since the paper's own standard is falsifiability (b576), I recommend adding per-proposal decisive comparisons, for example the same policy family with and without the streaming encoder at equal total compute for the camera-rate claim, and forecast-error decay over episodes with frozen policy weights for the single-attempt recalibration claim (b561), so that the central design thesis becomes testable as a whole rather than only module by module.","section":"Abstract; §5.3.3","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The sentence \"How cleanly the two streams divide recognition from control is contested [339], but that debate is not discussed in this paper\" is inconsistent with the text that follows: Section 4.2.2 does discuss the debate (\"how separate the channels are is debated\") and notes that the streams \"exchange object information rather than working apart\". The sentence reads like a leftover editorial note and should be removed or replaced with a pointer to the Section 4.2.2 caveat.","section":"§4.2.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Conclusion's claim that the self-labeling predictor's link \"is causal\" is too strong as stated: a map from command to outcome learned from the robot's own rollouts is causally interpretable only if the predictor conditions on all confounders of action and outcome, and state-dependent action selection can bias self-labeled data otherwise. A formulation such as \"predictive under the robot's own policy\" would be safer, or the conditioning assumption should be stated.","section":"§7 (Conclusion)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The term \"recursive self-improvement\" is used in the Abstract and again in the Conclusion but never defined; a one-sentence definition (for example, improvement that compounds without human intervention) would clarify the scope of the claim, especially since Section 6.1.5 limits today's systems to partial automation.","section":"Abstract; §7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Several claims that carry weight in the argument rest on non-archival or vendor sources: [109] (Figure AI Helix company report), [101] (Dyna-2 technical blog), and [62] (project page), alongside many 2026-dated preprints. This is understandable for a survey of a fast-moving area, but the text or reference list should mark vendor and preprint sources so that readers can weigh claims such as the 22% zero-shot success figure (b396).","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The correction search of Section 6.3.7 scores candidate corrections with the predictor (Section 5.3.1), but the paper does not discuss the early-deployment regime in which that predictor is still badly miscalibrated, where the search would follow the wrong direction; a brief discussion of this bootstrap problem (for example, using boundary-test outcomes as ground truth until the forecast error decays) would close the loop.","section":"§6.3.7; §5.3.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a serious position paper, and the stress-test concern about the judge is right but not disqualifying. The paper's actual contribution is a concrete blueprint—command stiffness alongside position, a dedicated streaming encoder for control next to the recognition encoder, a self-labeling predictor beside the policy, primitives verified at sensory boundaries, and practice/consolidation agents that move skills from code into weights. The combination is new even though the parts are borrowed, and the authors are unusually disciplined about labeling what is contested in the neuroscience and what is their interpretation. It reads like a survey that knows the difference between a result and a proposal.\n\nThe main soft spot is the third leg of the thesis: keeping improving after deployment. That leg requires an autonomous judge: attribution of failures to the right unit, scoring which outcome is closer to the goal, and verifying that a weight update is a real gain. The paper says in Section 6.1.5 that these judgments still fall to people, and Section 6.3.1 concedes that current models rarely diagnose contact failures. The practice and consolidation agents do not supply the missing capability; they restate it as a job description. The stress-test note lands. It is a genuine gap in the proposal, though an honestly acknowledged one, and the authors offer the whole design as hypotheses to be tested rather than as demonstrated results.\n\nSmaller issues: the conclusion at b561 reads like a result (\"the result is a robot that recalibrates...\") for something no one has built. That wording should be toned down. I would not make much of the preprint-heavy citation base; the quantitative claims are attributed to specific sources, and the paper flags the figures it treats as order-of-magnitude. The reference list is extensive and mostly appropriate, though I could not check completeness beyond the truncation point.\n\nWho gets value: anyone designing VLA architectures, compliant manipulation, tactile policies, or continual learning for robots. The survey maps the design space well, and each proposal is falsifiable by construction. It is not a results paper and should not be read as one. It deserves a serious referee; the right request is to keep the honest framing, tighten the conclusion, and state the judge problem as an explicit open problem with candidate mechanisms rather than naming an agent that assumes it.\n\nRecommendation: send to peer review. If I were the editor, I would let it through and ask for revision rather than desk-rejecting it.","headline":"A serious, honestly framed position paper whose self-improvement leg rests on a judge problem the authors name but do not solve.","tokens_in":57511,"tokens_out":2616,"duration_ms":19685},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Robots trained on a million trajectories still fail at everyday tasks a child masters; this survey argues the gap is design, not scale, and proposes a brain-inspired architecture to close it.","keywords":["robot learning","embodied AI","neuroscience","biological priors","vision-language-action models","world models","compliance control","motor primitives"],"falsifier":"A direct experiment would settle the central claim: hold the training data and inference budget fixed, and run the same manipulation suite — unfamiliar objects, scenes that change mid-motion, repeated trials — on a standard end-to-end VLA and on a system built with the proposed design: admittance compliance reading a tactile force field, a streaming control encoder, a self-labeling predictor, and primitive composition with boundary tests. The design-gap thesis predicts the designed robot wins at equal data; it is falsified if the standard VLA matches it once its data budget is scaled even tenfold. A sharper sub-test: if a control encoder distilled to predict velocity, optical flow, and time-to-contact does not beat an appearance-reconstruction encoder on camera-rate reaction tasks, the two-stream vision proposal fails on its own terms.","tokens_in":57928,"feed_emoji":"🧠","tokens_out":12839,"duration_ms":70130,"temperature":0.7,"pith_summary":"The paper argues that robot policies keep failing at everyday manipulation not because they lack data but because they are built the wrong way: one slow, recognition-shaped network reads a stale frame and emits precise positions into a fixed servo stack, is trained once, and never learns again after deployment. It proposes a next-generation robot model organized like the sensorimotor nervous system: compliance beneath position output, a dedicated fast vision encoder for control beside the slow recognition encoder, a self-labeling predictor beside the policy, action units composed from reusable primitives and verified at sensory boundaries, and practice and consolidation agents that gradually move proven skills from explicit code into network weights. If the design is right, a robot with this organization would solve unfamiliar tasks on first contact, react at camera rate to a moving world, and keep improving after deployment, using data it generates itself. The authors offer each proposal as a hypothesis to be tested rather than a demonstrated result.","feed_headline":"Robot dexterity gap is a design gap, not a scale gap","feed_subtitle":"Compliance, a fast vision stream, a self-teaching predictor, and practice: an agenda for robots that act on their own.","key_machinery":"The carrying mechanism is an architectural analogy, staked on a specific reading of what the sensorimotor brain does, in which each functional division maps to a concrete engineering component. Opposing muscle pairs set torque and stiffness independently, so the policy emits stiffness alongside position, realized by impedance or admittance control beneath the policy. Two visual streams motivate a streaming control encoder that precomputes motion, flow, and time-to-contact for a reactive head, separate from the recognition encoder. The cerebellum, read as a forward model, motivates a predictor beside the policy that is shallow, wide, and sparse, so an error-scaled test-time step corrects only the forecast that was wrong and leaves all others intact. Motor cortex and the prefrontal-to-motor hierarchy motivate primitive composition with verifiable goal tests at each action-unit boundary, plus practice and consolidation agents that move primitives from code into weights; body-part embeddings tie the modules together so the sensorimotor map transfers across embodiments. What this machinery does is replace one monolithic policy with a set of cooperating loops, each tuned to a different timescale, so that speed, generalization, and learning are handled by different components rather than by one network.","core_discovery":"The central claim, stated on the paper's own terms, is that the gap between a seven-year-old and a policy pretrained on over a million trajectories is a design gap: the brain's motor system solves speed, generalization, and lifelong improvement by dividing labor, and robotics has not yet copied those divisions. The blueprint has six moving parts. The policy should command stiffness alongside position, with touch as a force field telling the controller where to yield and how hard to press, so the world supplies much of the trajectory. A second, fast vision encoder, fed by motion-oriented signals distilled from video models, should drive a reactive policy head at camera rate while the slow encoder keeps recognition. A shallow, wide, sparse predictor beside the policy should forecast the sensory consequence of each command and recalibrate from its own error within a single episode. One model should compose action units from parameterized primitives, verify each at a sensory boundary, and mature each primitive from measuring code to typed function to end-to-end weights, guided by a practice agent that rehearses the weakest unit and a consolidation agent that graduates what works. The authors claim that together these features yield a robot that acts on its own initiative, which they call the basis of deployment and recursive self-improvement, and they state plainly that the neuroscience behind the analogy is partly hypothesis and that each transfer must be tested on real robots.","pith_inferences":["If the design-gap thesis is right, it yields a measurable systems-level prediction: beyond some data budget, these architectural changes should buy larger gains per unit of data than adding more pooled trajectories, a comparison that could be run on any standard manipulation benchmark.","The code-to-weights path inverts today's data economy: code-writing frontier models become the teachers that generate the measured parameters and correction records used to train fast reactive heads, so the paper's own concession that current models rarely write correct measuring code or diagnose contact failures is the true gating bottleneck of the whole agenda.","The two-stream vision proposal suggests a concrete distillation recipe worth testing: distill the control encoder to predict velocity, optical flow, and time-to-contact rather than to reconstruct appearance, then compare reaction speed and contact-rich success at equal compute against a recognition-pretrained encoder.","If consolidation works as described, the durable asset of a robot fleet becomes the shared primitive library rather than per-robot weights, implying that cross-embodiment transfer and skill reuse should be organized around the library's growth — an economic argument the paper does not make."],"forward_implications":["A policy that commands stiffness alongside a position target would let contact-rich skills run on the position-controlled arms most robots use, because touch rather than precise trajectory computation would decide where to yield and how hard to press.","A camera-rate control encoder would remove the blindness inside action chunks: a robot could correct a slipping grasp or a moved target mid-motion instead of committing to an executed sequence it cannot see.","A self-labeling predictor that recalibrates from its own error would let a robot adapt to a stiffer lock or a heavier payload within a single attempt while leaving the policy's weights untouched, avoiding the catastrophic forgetting that naive fine-tuning risks.","Composing new tasks from verified primitives would let a robot face an unfamiliar task on the first try, and the practice-plus-consolidation loop would let it keep improving after deployment rather than freezing at the factory door.","Because every executed action labels its own outcome a moment later, most of the training signal this design needs — tactile consequences, forecast errors, boundary-test verdicts — is available in ordinary deployment data with no human annotation."],"supporting_citations":[{"why":"The flow-matching VLA is the representative current architecture whose speed, mid-chunk blindness, and frozen weights the proposals are designed to fix.","marker":"[42]"},{"why":"The action-as-text VLA baseline whose generalization limits and training stages the survey argues against.","marker":"[210]"},{"why":"Code as Policies establishes that language models can write manipulation programs, the starting point of the code-to-weights consolidation path.","marker":"[239]"},{"why":"Establishes action chunking, the mechanism whose blindness between chunks motivates the streaming encoder and reactive policy head.","marker":"[450]"},{"why":"Supplies the two-visual-streams division of primate vision that motivates the dedicated fast control encoder beside the recognition encoder.","marker":"[129]"},{"why":"The forward-model reading of the cerebellum that motivates the self-labeling predictor running beside the policy.","marker":"[412]"},{"why":"The Smith-predictor structure that the proposed predictor design borrows for covering sensing delay.","marker":"[269]"},{"why":"The precision-grip measurements showing grip force rises about 74 ms after a micro-slip, anchoring the tactile compliance and reflex proposals.","marker":"[189]"},{"why":"Evidence that stimulating motor cortex evokes whole actions, the empirical basis for the claim that motor cortex holds reusable action primitives.","marker":"[138]"},{"why":"The attenuation of self-generated touch (why one cannot tickle oneself) supports prediction of sensory consequences as a mechanism the robot predictor instantiates.","marker":"[44]"}],"fun_headline_variants":["Robot dexterity needs the brain's division of labor","From cerebellum to controller: a robot learning agenda","The gap is design: what robot policies should copy from the brain","Compliance, fast vision, and practice: the borrowed robot brain","Robot learning should copy the sensorimotor brain's playbook"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The blueprint stands on the premise that the organizational features it exports from the brain — separate fast vision for action, a predictive cerebellum, motor-cortex primitives with boundary checks, and offline consolidation — are the load-bearing causes of human speed, generalization, and self-improvement rather than biological side effects, and that they survive transplantation into an engineered robot; the paper acknowledges both the debate in neuroscience and that the premise is testable only on real hardware, and it additionally relies on frontier multimodal models becoming able to write correct measuring code and diagnose contact failures, a capability it concedes current models rarely have.","fun_headline_variants_meta":{"raw":{"variants":["Robot dexterity needs the brain's division of labor","From cerebellum to controller: a robot learning agenda","The gap is design: what robot policies should copy from the brain","Compliance, fast vision, and practice: the borrowed robot brain","Robot learning should copy the sensorimotor brain's playbook"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"cost":0.000295568,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000295568,"upstream_inference_prompt_cost":3.4944e-05,"upstream_inference_completions_cost":0.000260624},"total_tokens":2632,"prompt_tokens":1200,"completion_tokens":1432,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":1020,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1361}},"tokens_in":1200,"tokens_out":1432,"cost_source":"openrouter_usage","duration_ms":11353,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":1361,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-03T21:17:27.520969+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A direct experiment would settle the central claim: hold the training data and inference budget fixed, and run the same manipulation suite — unfamiliar objects, scenes that change mid-motion, repeated trials — on a standard end-to-end VLA and on a system built with the proposed design: admittance compliance reading a tactile force field, a streaming control encoder, a self-labeling predictor, and primitive composition with boundary tests. The design-gap thesis predicts the designed robot wins at equal data; it is falsified if the standard VLA matches it once its data budget is scaled even tenfold. A sharper sub-test: if a control encoder distilled to predict velocity, optical flow, and time-to-contact does not beat an appearance-reconstruction encoder on camera-rate reaction tasks, the two-stream vision proposal fails on its own terms.","supporting_citations":[{"cited_title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","cited_arxiv_id":"2410.24164","evidence_quote":"The flow-matching VLA is the representative current architecture whose speed, mid-chunk blindness, and frozen weights the proposals are designed to fix."},{"cited_title":"OpenVLA: An Open-Source Vision-Language-Action Model","cited_arxiv_id":"2406.09246","evidence_quote":"The action-as-text VLA baseline whose generalization limits and training stages the survey argues against."},{"cited_title":"Code as Policies: Language Model Programs for Embodied Control,","cited_arxiv_id":null,"evidence_quote":"Code as Policies establishes that language models can write manipulation programs, the starting point of the code-to-weights consolidation path."},{"cited_title":"Goodale and A","cited_arxiv_id":null,"evidence_quote":"Supplies the two-visual-streams division of primate vision that motivates the dedicated fast control encoder beside the recognition encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Smith-predictor structure that the proposed predictor design borrows for covering sensing delay."},{"cited_title":"Johansson and Göran Westling","cited_arxiv_id":null,"evidence_quote":"The precision-grip measurements showing grip force rises about 74 ms after a micro-slip, anchoring the tactile compliance and reflex proposals."},{"cited_title":"Complexmovementsevokedbymicrostimulation of precentral cortex.Neuron, 34(5):841–851, 2002","cited_arxiv_id":null,"evidence_quote":"Evidence that stimulating motor cortex evokes whole actions, the empirical basis for the claim that motor cortex holds reusable action primitives."},{"cited_title":"Wolpert, and Chris D","cited_arxiv_id":null,"evidence_quote":"The attenuation of self-generated touch (why one cannot tickle oneself) supports prediction of sensory consequences as a mechanism the robot predictor instantiates."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}