{"id":"804902eb-68ae-40db-930c-55a4ca3b1d0a","arxiv_id":"2606.31941","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"LeCropFollow enables zero-shot sim-to-real visual navigation in unstructured crop fields by optimizing trajectories in a latent manifold from self-supervised semantic heatmaps with TD-MPC2, reducing semantic failures 2.4x versus keypoint baselines in corn gaps.","lead":"LeCropFollow is a navigation system for farm robots that plans paths in a learned latent space from semantic image heatmaps using model-based reinforcement learning instead of building explicit maps. This could let robots handle irregular rows and gaps in real fields where geometric methods often fail.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Self-supervised heatmap extractor may not reliably preserve uncertainty for latent optimization in gaps","rationale":"The reader's weakest_assumption matches the single point whose failure would falsify the central representational claim. Because the abstract supplies no quantitative evidence (ablation tables, uncertainty metrics, or sim-to-real distribution shift measurements) that the extractor actually retains the required information, the concern remains load-bearing even after reading the full manuscript description.","tokens_in":1742,"tokens_out":312,"duration_ms":19260,"concrete_test":"Re-run the field gap trials with an otherwise identical TD-MPC2 planner whose input is replaced by a deterministic keypoint or depth-derived map (same training regime, same reward); if the semantic-failure rate rises to within 20% of the keypoint baseline, the latent-heatmap representation is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline performance gain (2.4x fewer semantic failures in plantation gaps, zero-shot sim-to-real) rests on the claim that operating directly on the uncompressed self-supervised heatmap inside TD-MPC2 supplies the semantic context and uncertainty that geometric/keypoint methods discard. If the extractor—trained self-supervised—collapses or misrepresents uncertainty in visually ambiguous gaps, the latent-space advantage disappears and the observed improvement could be attributable to other unablated factors (planner hyperparameters, reward shaping, or simulation fidelity). No section in the provided text supplies an ablation that isolates the heatmap representation from the rest of the pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents LeCropFollow, a visual navigation framework for unstructured crop fields that integrates a self-supervised semantic heatmap extractor with the TD-MPC2 MBRL planner to perform trajectory optimization directly in latent space. It claims this approach preserves semantic context and uncertainty discarded by geometric methods, enabling zero-shot sim-to-real transfer and superior performance in plantation gaps (2.4x reduction in semantic failures vs. keypoint baselines) while matching SOTA in unstructured rows.","tokens_in":1858,"tokens_out":310,"duration_ms":26815,"significance":"If the empirical claims hold after proper validation, the work would be significant for agricultural robotics, as it provides evidence that latent-space planning over uncompressed self-supervised representations can outperform geometric reductions in ambiguous, unstructured terrain without requiring fine-tuning.","major_comments":[{"comment":"Abstract: the central performance claims (2.4x reduction in semantic failures, zero-shot transfer) are stated without any description of experimental design, number of trials, error bars, data exclusion rules, or statistical tests, rendering the headline result unverifiable from the supplied information.","section":"Abstract"},{"comment":"No ablation isolating the self-supervised heatmap extractor (and its claimed preservation of uncertainty) from planner hyperparameters, reward shaping, or simulation details is provided; without this, it is impossible to attribute the gap-specific gains to the latent representation rather than other unablated factors.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on the clarity of our claims and the need for stronger isolation of contributions. We address each major comment below.","responses":[{"response":"We agree that the abstract is overly concise and does not convey the experimental context. The manuscript body (Experiments section) describes the field trials, metrics, and comparisons, but the abstract does not. In the revision we will expand the abstract with a brief clause summarizing the experimental design and validation approach to make the headline claims verifiable at a glance.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central performance claims (2.4x reduction in semantic failures, zero-shot transfer) are stated without any description of experimental design, number of trials, error bars, data exclusion rules, or statistical tests, rendering the headline result unverifiable from the supplied information."},{"response":"We acknowledge this limitation. The current results compare the full LeCropFollow system against keypoint baselines but do not hold the TD-MPC2 planner, reward, and simulation fixed while swapping only the representation. We will add a controlled ablation in the revision that replaces the self-supervised heatmap input with geometric features inside the same planner to better isolate the contribution of the latent representation.","revision_made":"yes","referee_comment":"No ablation isolating the self-supervised heatmap extractor (and its claimed preservation of uncertainty) from planner hyperparameters, reward shaping, or simulation details is provided; without this, it is impossible to attribute the gap-specific gains to the latent representation rather than other unablated factors."}],"tokens_in":1334,"tokens_out":353,"duration_ms":25153,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main move is to skip geometric compression and instead run TD-MPC2 directly on uncompressed self-supervised semantic heatmaps for under-canopy navigation. It reports matching standard baselines in regular rows while cutting semantic failures by 2.4x in plantation gaps, plus zero-shot sim-to-real without fine-tuning. That combination for this specific setting looks like the actual new piece.\n\nThe framing is clear: geometric methods throw away uncertainty and context that matter in irregular terrain. Running the planner on the raw heatmap signal is a reasonable way to keep that information. Field tests in late-stage corn are the right kind of evidence for an ag-robotics paper, and releasing code, models, and data is useful.\n\nThe soft spots are straightforward. The abstract states the 2.4x gain and the zero-shot result but gives no numbers on trial count, variance, exclusion criteria, or statistical tests. More importantly, there is no sign of an ablation that isolates the heatmap representation from planner hyperparameters, reward design, or simulation details. The stress-test concern holds: if the extractor does not reliably carry uncertainty through ambiguous gaps, the claimed advantage could come from elsewhere. Without those checks, the central performance claim stays hard to pin down.\n\nThis is for people working on field robotics or latent MBRL in unstructured settings. A reader who needs practical navigation ideas in crops will get something from the results and the released assets. It deserves a serious referee because the problem is real, the field data exist, and the approach is concrete, even if the current write-up leaves the attribution questions open.","headline":"LeCropFollow shows a latent planning setup with self-supervised heatmaps and TD-MPC2 that claims 2.4x fewer failures in crop gaps and zero-shot transfer, but the abstract supplies no experimental details or ablations to back the attribution.","tokens_in":2364,"tokens_out":417,"would_cite":false,"duration_ms":18001,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Navigation through irregular crop gaps succeeds when trajectories are optimized directly over semantic heatmaps in latent space rather than reduced to geometric points.","keywords":["visual navigation","latent space planning","agricultural robotics","unstructured environments","semantic heatmaps","model-based reinforcement learning","zero-shot transfer","crop field navigation"],"falsifier":"Physical field trials in which LeCropFollow produces more semantic failures than keypoint-based baselines while crossing plantation gaps would falsify the performance advantage.","tokens_in":2657,"feed_emoji":"🌽","tokens_out":669,"duration_ms":27862,"temperature":0.7,"pith_summary":"The paper argues that geometric compression of visual data throws away the semantic context and uncertainty required to cross ambiguous terrain such as plantation gaps. LeCropFollow replaces that compression with a learned latent representation built from a self-supervised semantic heatmap extractor, then plans trajectories inside that manifold using model-based reinforcement learning. Because the full heatmap signal is retained, the system transfers from simplified simulation to physical late-stage corn fields without any fine-tuning. Field tests show the method matches existing approaches along regular rows yet produces 2.4 times fewer semantic failures when rows are interrupted. A reader would care because agricultural robots must operate reliably in heterogeneous, under-canopy conditions where perfect maps or explicit geometry are unavailable.","feed_headline":"Latent planning cuts crop robot failures 2.4x in gaps","feed_subtitle":"LeCropFollow keeps semantic context that geometric methods discard and transfers zero-shot from simulation to real late-stage corn fields.","key_machinery":"A self-supervised semantic heatmap extractor whose output is fed directly to a model-based reinforcement learning planner that optimizes trajectories inside the learned latent manifold.","core_discovery":"LeCropFollow shows that operating over uncompressed semantic heatmaps inside a latent manifold, rather than compressing visuals into deterministic spatial references, enables robust navigation through irregular planting and discontinuities; the resulting zero-shot sim-to-real deployment matches state-of-the-art performance in structured rows and delivers a 2.4 times reduction in semantic failures relative to keypoint-based methods inside plantation gaps.","pith_inferences":["The same latent-heatmap approach could be tested in other unstructured outdoor settings such as orchards or forest understories where row discontinuities also appear.","If the heatmap signal already encodes uncertainty, the method may reduce the need for additional uncertainty-aware sensors or explicit mapping passes.","Replacing the current planner with alternative latent-space optimizers would offer a direct test of whether the representational choice, rather than any single planner, drives the observed robustness."],"forward_implications":["Zero-shot transfer from simplified simulation to physical corn fields occurs without any fine-tuning.","Performance equals existing geometric methods along continuous unstructured rows.","Semantic failures drop by a factor of 2.4 relative to keypoint methods inside plantation gaps.","Latent planning supplies a practical alternative to geometric estimation for heterogeneous agricultural settings."],"fun_headline_variants":["LeCropFollow plans in latent space for crop field navigation","Latent planning handles unstructured plantation gaps","LeCropFollow shows 2.4x fewer failures in gaps","Latent manifold supports zero-shot corn field transfer"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The self-supervised semantic heatmap extractor must capture and retain enough semantic context and uncertainty from raw images to support reliable trajectory optimization in the latent space when terrain is ambiguous.","fun_headline_variants_meta":{"raw":{"variants":["LeCropFollow plans in latent space for crop field navigation","Latent planning handles unstructured plantation gaps","LeCropFollow shows 2.4x fewer failures in gaps","Latent manifold supports zero-shot corn field transfer"]},"model":"grok-4.3","cost_usd":0.00917,"raw_usage":{"total_tokens":4113,"prompt_tokens":675,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":91699500,"prompt_tokens_details":{"text_tokens":675,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3377,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":675,"tokens_out":61,"duration_ms":27396,"temperature":1.0,"reasoning_tokens":3377,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T05:14:14.170203+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Physical field trials in which LeCropFollow produces more semantic failures than keypoint-based baselines while crossing plantation gaps would falsify the performance advantage.","supporting_citations":[],"review_version":1}