{"id":"f3855faa-6446-45b2-a220-6f81fbede8bf","arxiv_id":"2606.09183","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A burial-conditioned curriculum in particle simulation trains effective policies for excavator obstacle removal that transfer successfully to a real 12-ton excavator.","lead":"The paper describes a curriculum learning approach using particle simulation to train policies for excavators to remove buried obstacles, with successful transfer to a real 12-ton machine. A smart generalist might read it to see how simulation-based training with progressive difficulty can enable autonomous operation in variable real-world conditions like construction sites.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Particle simulation fidelity for real soil-obstacle interactions remains the least-secured precondition for the sim-to-real transfer claim.","rationale":"The reader's weakest assumption directly identifies the same load-bearing precondition. The full manuscript text does not supply independent evidence (such as sim-real matching experiments) that would remove this dependency, so the provisional UNVERDICTED status is unaffected.","tokens_in":1796,"tokens_out":339,"duration_ms":24665,"concrete_test":"Re-run the curriculum training in the particle simulator but with an independent soil model (e.g., DEM with different friction parameters or a mesh-based alternative) at the same burial depths; if the resulting policy fails to transfer at the same real-world success rate reported in the paper, the original simulator's fidelity was load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that policies trained exclusively in the particle simulator (observing RGB-D terrain/obstacle states and outputting parameterized trajectories) produce behavior that succeeds on a physical 12-ton excavator. This holds only if the simulator accurately reproduces the state-dependent dynamics—contact states, terrain deformation, and obstacle visibility—across the controllable burial depths used in the curriculum. The curriculum explicitly scales particle count with burial depth to manage cost while increasing difficulty; any mismatch in how particles model real soil flow or burial resistance at deeper levels would mean the learned policy exploits simulator-specific artifacts. The abstract reports real-world success, yet the argument provides no quantitative sim-to-real validation (e.g., side-by-side deformation maps or contact metrics) to confirm the simulator is not the source of the performance gap versus baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a burial-conditioned curriculum learning framework for training excavation policies in particle-based simulation. The policy takes RGB-D observations of terrain and obstacles and outputs parameterized trajectories. The curriculum starts with shallow burials (low particle count, low difficulty) and progressively increases burial depth while scaling particle count to control both task difficulty and simulation cost. Experiments claim the approach learns an effective policy in three days of training, while baselines fail after a full week, and that the resulting policy transfers successfully to a physical 12-ton excavator performing obstacle removal on open ground with steel obstacles.","tokens_in":1945,"tokens_out":553,"duration_ms":15102,"significance":"If the sim-to-real transfer claim holds, the work demonstrates a practical route to learning state-dependent excavation policies that adapt to changing soil-obstacle conditions, a long-standing barrier in autonomous earthmoving. The curriculum that jointly modulates burial depth and particle count is a concrete engineering contribution that addresses the computational cost of repeated-cycle particle simulation. The manuscript supplies no machine-checked proofs or parameter-free derivations, but the reproducible observation-action interface between simulator and real excavator is a strength that supports the transfer narrative.","major_comments":[{"comment":"Abstract and Experiments section: The central claim that the curriculum 'successfully learns an effective obstacle-removal policy' and 'achieves successful transfer' is presented without any quantitative metrics, success rates, error bars, statistical tests, or definition of how success was measured (e.g., number of cycles to removal, failure modes). This absence prevents evaluation of the performance gap versus baselines.","section":"Abstract / Experiments"},{"comment":"Simulation and Transfer sections: The sim-to-real claim rests on the unvalidated assumption that the particle simulator reproduces real soil-obstacle contact states, terrain deformation, and obstacle visibility across the burial depths used in the curriculum. No side-by-side quantitative validation (deformation maps, contact force metrics, or burial-resistance curves) is reported; any systematic mismatch would mean the policy exploits simulator artifacts rather than learning transferable behavior.","section":"Simulation / Transfer"}],"minor_comments":[{"comment":"The abstract states that 'baseline methods fail even after a full week of training' without specifying the baseline algorithms, their hyper-parameters, or training details; this information belongs in the Experiments section for reproducibility.","section":"Abstract"},{"comment":"Notation for the parameterized trajectory (e.g., what parameters are output by the policy) is introduced without an explicit equation or diagram in the methods description.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to incorporate quantitative metrics and clarify sim-to-real aspects where possible.","responses":[{"response":"We agree that the abstract and experiments lack explicit quantitative metrics. In the revised version we will add success rates (percentage of successful removals over repeated trials), definition of success (e.g., obstacle fully removed within a maximum number of cycles), failure modes, error bars from multiple random seeds, and statistical comparisons against baselines. These additions will appear in both the abstract and the experiments section.","revision_made":"yes","referee_comment":"[Abstract / Experiments] Abstract and Experiments section: The central claim that the curriculum 'successfully learns an effective obstacle-removal policy' and 'achieves successful transfer' is presented without any quantitative metrics, success rates, error bars, statistical tests, or definition of how success was measured (e.g., number of cycles to removal, failure modes). This absence prevents evaluation of the performance gap versus baselines."},{"response":"The manuscript reports no quantitative side-by-side validation of simulator outputs against real soil measurements. We will revise the relevant sections to explicitly state this limitation, describe the reproducible observation-action interface used for transfer, and present the real-world excavator results as the primary empirical evidence of transfer. We will also add any available qualitative comparisons if they exist in our data.","revision_made":"partial","referee_comment":"[Simulation / Transfer] Simulation and Transfer sections: The sim-to-real claim rests on the unvalidated assumption that the particle simulator reproduces real soil-obstacle contact states, terrain deformation, and obstacle visibility across the burial depths used in the curriculum. No side-by-side quantitative validation (deformation maps, contact force metrics, or burial-resistance curves) is reported; any systematic mismatch would mean the policy exploits simulator artifacts rather than learning transferable behavior."}],"tokens_in":1476,"tokens_out":417,"duration_ms":17827,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main things to know are that the authors use a curriculum that starts with shallow burial and ramps up depth while cutting particle count to control both difficulty and compute cost, and that the resulting policy works on a physical excavator where baselines do not.\n\nWhat is new is the explicit link between burial condition and simulation expense in the curriculum. The paper does a reasonable job laying out the observation-action interface (RGB-D to parameterized trajectories) and showing that the same setup runs in sim and on the real 12-ton machine with steel obstacles.\n\nThe soft spots are clear from the abstract. There are no success rates, no error bars, no statistical tests, and no description of how success was scored or how the baselines were trained. The bigger issue is the missing check on particle-sim fidelity. The stress-test note is right: without side-by-side comparisons of contact states, deformation, or obstacle visibility, it is hard to tell whether the policy learned real dynamics or just simulator quirks. The abstract does not address this.\n\nThis is for people working on construction robotics and sim-to-real transfer for heavy equipment. A reader who wants a concrete example of curriculum design for variable soil conditions would get something out of the real-world transfer result. The work is coherent enough on its own terms to deserve peer review, even if the reviewers will need to see the missing metrics and validation data.","headline":"The burial-depth curriculum lets them train an excavator policy in particle sim that transfers to a real 12-ton machine, but the abstract gives almost no numbers and no check on whether the sim matches real soil behavior.","tokens_in":2474,"tokens_out":367,"would_cite":false,"duration_ms":16389,"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":"A burial-depth curriculum in particle simulation trains an excavator policy for obstacle removal that transfers successfully to a real 12-ton machine.","keywords":["obstacle removal","excavator","policy learning","particle simulation","curriculum learning","sim-to-real","robotics","earthwork"],"falsifier":"Direct real-world trials in which the transferred policy fails to remove obstacles that the simulated policy removed under matching burial depths and obstacle placements would falsify the transfer claim.","tokens_in":2697,"feed_emoji":"🚜","tokens_out":494,"duration_ms":27254,"temperature":0.7,"pith_summary":"The paper aims to automate repeated excavation cycles for removing obstacles whose burial conditions change over time. It argues that particle-based simulation can capture the needed soil-obstacle contacts, terrain changes, and visibility, but only if computational cost is managed. The proposed approach uses a curriculum that begins with shallow burials and steadily increases depth while lowering particle count, thereby raising difficulty and lowering simulation expense in tandem. The resulting policy takes RGB-D measurements as input and produces parameterized trajectories as output, using the identical interface in simulation and on the physical excavator. A sympathetic reader would care because the method shows how to make long-horizon robotic earthmoving policies trainable at all.","feed_headline":"Three-day curriculum teaches excavator to clear buried obstacles","feed_subtitle":"Particle simulation with progressive burial depth trains a policy that transfers to a real 12-ton machine while baselines fail after a week.","key_machinery":"The burial-conditioned curriculum that varies obstacle burial depth to control both task difficulty and the number of particles required for accurate simulation.","core_discovery":"The paper claims that a time-efficient sim-to-real policy learning framework using particle simulation under controllable burial conditions, together with a curriculum that progressively increases burial depth while adjusting particle count, produces an effective obstacle-removal policy within three days. The policy observes terrain and obstacle information from RGB-D measurements and outputs a parameterized excavation trajectory. Experiments show this policy transfers to a real 12-ton excavator operating on open ground with various steel obstacles, while baseline methods fail even after a full week of training.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Burial-conditioned curriculum trains policy in three days with particle simulation","Policy learns from RGB-D to remove obstacles after three-day sim curriculum","Adjusts particle count as burial deepens for efficient three-day training","Baselines fail after full week as curriculum succeeds in three days"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The particle simulation must accurately reproduce real-world soil-obstacle interactions including contact states, terrain deformation, and obstacle visibility under the burial conditions used in training.","fun_headline_variants_meta":{"raw":{"variants":["Burial-conditioned curriculum trains policy in three days with particle simulation","Policy learns from RGB-D to remove obstacles after three-day sim curriculum","Adjusts particle count as burial deepens for efficient three-day training","Baselines fail after full week as curriculum succeeds in three days"]},"model":"grok-4.3","cost_usd":0.006152,"raw_usage":{"total_tokens":2936,"prompt_tokens":735,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":61524500,"prompt_tokens_details":{"text_tokens":735,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2130,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":735,"tokens_out":71,"duration_ms":12462,"temperature":1.0,"reasoning_tokens":2130,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T16:32:27.673758+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Direct real-world trials in which the transferred policy fails to remove obstacles that the simulated policy removed under matching burial depths and obstacle placements would falsify the transfer claim.","supporting_citations":[],"review_version":1}