{"id":"f1489fe5-4397-4344-b54d-939a144aab9d","arxiv_id":"2411.12711","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UBSoft is a simulation platform that uses robot-centered adaptive resolution to make unbounded soft-environment robot skill learning tractable, with benchmark tasks and sim-to-real demos.","lead":"The authors built a simulation platform called UBSoft that lets robots practice skills in very large sand, snow, and other soft environments by running detailed physics only near the robot and cheaper approximations farther away. This matters because robot learning in soft materials was previously limited to small, bounded boxes, and the platform could unlock training for locomotion and manipulation in open, realistic terrain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Split/merge resampling in Sec 3.2 is not specified or validated; the single falling-cube test does not exercise repeated hierarchy traversal, so the efficiency-accuracy tradeoff is unproven for long-horizon scenes.","rationale":"The reader's weakest assumption already flags that 'particles are handled correctly as the moving hierarchy traverses the scene' and that the falling-cube test is insufficient. My stress-test focuses on one concrete mechanism behind that assumption: the split/merge rules for deformation-gradient-carrying particles are unspecified and unvalidated. This is the most load-bearing point because if split/merge corrupts plastic state, the entire efficiency-accuracy tradeoff collapses, even before considering boundary treatment. I do not move the verdict because the reader's CONDITIONAL verdict already captures this uncertainty; the paper remains a plausible platform contribution pending code release and additional validation. The real-world transfer and benchmark suite are useful evidence, but they do not isolate the resampling error.","tokens_in":13945,"tokens_out":6555,"duration_ms":73009,"concrete_test":"Run the Quadruped Sand Walk task for 5000 simulation steps under both UBSoft and uniform MLS-MPM 256, logging per-step Chamfer distance of sand particle sets and cumulative foot contact forces, while instrumenting every split/merge event. If Chamfer error grows with repeated traversals, or if any single merge/split event causes a contact-force deviation above 5% of the peak force, the resampling scheme does not preserve dynamics and the accuracy claim is not supported for long-horizon scenes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that spatially adaptive MLS-MPM preserves dynamics while saving an order of magnitude. The load-bearing mechanism is the particle split/merge operation in Sec 3.2 and App. A. For elastoplastic materials modeled with multiplicative F = F_e F_p (snow, sand), resampling requires partitioning/combining deformation gradients and plastic history. The paper states only that split conserves momentum and merge averages 'other maintained information' (Sec 3.2), with no equations or validation. Averaging total F across particles with different strain histories is not equivalent to coarse-graining an elastoplastic continuum; it can erase plastic strain, inject spurious stress, or violate the F_e/F_p decomposition. Moreover, in locomotion tasks feet repeatedly enter/exit fine regions, so sand under the robot is split and merged every step; errors may accumulate. Table 1's falling-cube benchmark provides only a single Chamfer distance at the end state, no error trajectory, no repeated traversal, and no force/velocity comparison; it cannot distinguish discretization error from resampling artifacts. The 'unbounded' boundary treatment is a related unstated assumption: particles leaving the outermost grid are never described, yet long-horizon walking relies on them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces UBSoft, a simulation platform for robot skill learning in large-scale soft-body environments. The core contribution is a spatially adaptive MLS-MPM engine that uses hierarchical grids centered on the robot agent, with finer grids and smaller particles near the agent and coarser representations farther away. The platform is claimed to reduce storage and computation by over an order of magnitude while preserving simulation fidelity. The paper also presents eight benchmark tasks (four manipulation, four locomotion), evaluates PPO, SAC, CMA-ES, and differentiable-physics trajectory optimization, and reports open-loop sim-to-real transfer for two manipulation tasks on an XArm.","tokens_in":14260,"tokens_out":4896,"duration_ms":49656,"significance":"If the spatially adaptive resampling mechanism is sound, UBSoft addresses a genuine bottleneck: existing MPM-based simulators are limited to small, bounded domains, whereas robot locomotion and manipulation in sand or snow require large scenes. The benchmark suite and fully differentiable implementation are potentially useful to the robot-learning community. However, the central resampling scheme is under-specified, the accuracy claims rest on a single scene and a single scalar metric, and the treatment of the \"unbounded\" boundary is not described. The paper's value depends on details that are currently deferred to code and appendix text, so the contribution is promising but not yet established.","major_comments":[{"comment":"The split/merge resampling is the load-bearing component of the adaptive scheme, but it is described only verbally: 'other maintained information on the particles is averaged according to specific rules' with no equations. For elastoplastic materials modeled with multiplicative F = Fe Fp (sand, snow, elastic), averaging the deformation gradient or other state without specifying how Fe/Fp and plastic history are transformed can erase plastic strain or inject spurious stress. Please provide explicit update rules for position, velocity, mass/volume, F, and plasticity state during split and merge, and validate the resampling on a scene that exercises repeated split/merge cycles (e.g., a robot walking a long distance), rather than the single falling-cube test in Sec. 5.1.","section":"Sec. 3.2 and Appendix A"},{"comment":"The efficiency-accuracy claim rests on one falling-cube scene, a single end-of-simulation Chamfer distance computed against an MLS-MPM 256 approximation that is not ground truth. There is no error trajectory over time, no repeated trials, no statistical uncertainty, and no comparison of contact forces or agent-relevant quantities such as foot penetration. The claim that the adaptive result 'closely matches' MLS-MPM 256 is therefore not established beyond a single sample. Additionally, the dotted-line speedup/storage curves in Fig. 4 are extrapolated ('transitional MPM fails to simulate'), so the order-of-magnitude benefit for large scenes is asserted rather than measured.","section":"Sec. 5.1, Table 1 and Fig. 4"},{"comment":"The manuscript never specifies what happens to particles that leave the outermost hierarchical grid as the robot moves, nor how particles behind the agent are handled when the grid is re-centered. The paper's central claim of supporting 'unbounded' environments depends directly on this mechanism. For long-horizon locomotion, the robot traverses large distances and the moving grid boundary will inevitably intersect the soft material; without an explicit algorithm for particle transfer, outflow, or reinsertion, the unboundedness claim is unsupported. Please add a precise description of the grid-motion and boundary treatment.","section":"Sec. 3.2 and App. A"},{"comment":"The paper repeatedly refers to 'the code' ('Readers are recommended to refer to the code and Appendix A') but provides no code repository URL, supplementary material, or pseudocode. For a platform paper whose central contribution is the adaptive implementation, the missing resampling equations and hierarchical-grid stencil details mean the work cannot be independently reproduced or assessed. Make the code/implementation available or include the necessary details in the paper.","section":"Sec. 3.2 and reproducibility"},{"comment":"The sim-to-real transfer section is qualitative only: the text states that the robot writes 'CoRL' and scoops a cube, with no quantitative success rates, comparisons against a non-transferred baseline, or multiple trials. The conclusion that UBSoft has a 'small sim-to-real gap' is not supported by the reported evidence. Please add quantitative evaluation (e.g., success rate, trajectory tracking error, or sand-painting Chamfer distance in the real scene) or temper the claim accordingly.","section":"Sec. 5.3 and Appendix C.2"}],"minor_comments":[{"comment":"The platform name is rendered inconsistently as 'UBS OFT' in the abstract, introduction, and conclusion; it should be 'UBSoft' everywhere.","section":"Throughout"},{"comment":"The hierarchical-grid description defines li = 2^i l0 and says the grid count per dimension is '2 * k', but k is never defined in terms of layer count L or the grid extents. Define k and L explicitly.","section":"Sec. 3.2"},{"comment":"There is a typo in 'unbounded sandy pr snowy landscape'; 'pr' should be 'or'.","section":"Sec. 4.2"},{"comment":"The line 'We sample NS = PL l=1 Nl = 200 particles' is unclear about whether Nl is a per-layer count or whether the sum is over layers; specify how many particles are sampled per layer and how the total of 200 is obtained.","section":"Appendix B.1"},{"comment":"Some reward values appear suspiciously duplicated with different standard deviations (e.g., Quadruped Snow Walk PPO mean 2387.4 in Table 2 vs Quadruped Elastic Walk PPO mean 2387.4 in Table 3). Verify that these are not copy-paste errors.","section":"Table 3"},{"comment":"The comparison across methods is based on raw rewards that have different scales per task; reporting normalized success rates or per-task completion criteria would make the conclusions more robust.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is marked as a CoRL 2024 submission; if it is under consideration at a journal, the validation bar will be higher. The central missing element is the split/merge specification: the current text defers to 'the code' without a URL, which is problematic for a platform paper. The accuracy comparison in Table 1 is a single scene and a single metric, so the headline claims are weaker than they appear. The sim-to-real results are also qualitative. I recommend major revision with a request for the missing technical details and quantitative validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"UBSoft is a solid engineering contribution: a robot-centered hierarchical MPM with particle split/merge that makes large-scale soft terrain simulation tractable, plus a benchmark of eight tasks spanning manipulation and locomotion. The efficiency numbers look good, and the sim-to-real demos are a nice touch. But the load-bearing resampling step is under-specified and the accuracy validation is a single falling-cube test, so the platform is not yet proven for long-horizon locomotion where particles repeatedly cross resolution boundaries.\n\nWhat is genuinely new here is the specific combination—agent-centered nested grids with split/merge, fully differentiable, with rigid coupling and benchmark tasks—applied to robotics. The underlying adaptive MPM techniques exist in graphics, but the system integration and the task suite are a real contribution. The comparison to uniform-resolution MLS-MPM is honest: they don't claim to beat 256 uniformly; they match it while using an order of magnitude less storage and time on that test. The reward curves across PPO, SAC, CMA-ES, and DiffPhys give the community a useful baseline picture.\n\nThe main soft spot is the particle split/merge in Sec 3.2. The text says momentum is conserved and 'other maintained information' is averaged, with no equations. For Drucker-Prager sand and snow, the deformation gradient and plastic history matter; averaging total F across particles with different strain histories can erase plastic strain or inject spurious stress. This isn't a mathematical showstopper—a careful resampling scheme could work—but it needs to be specified and validated. The falling-cube benchmark gives one final Chamfer distance, no error trajectory, no force or velocity comparison, and no repeated traversal. In walking tasks, the feet split and merge sand every step, so errors could accumulate. The boundary treatment for particles leaving the outermost grid is also unstated, despite the 'unbounded' claim. And the code isn't released, which makes it hard to verify the implementation.\n\nThese are fixable issues, not fatal flaws. The central claim—that adaptive resolution saves an order of magnitude while staying close to a high-resolution reference—is plausible and partially supported. To make the platform widely trusted, I'd want to see the split/merge equations, a long-horizon validation with repeated boundary crossings, boundary handling, and code release.\n\nWho is this for? Researchers working on simulation for robot learning in deformable or granular media, and people building benchmarks. It deserves serious peer review with the expectation of revisions. I'd recommend accepting it conditionally and pushing for those details.","headline":"A useful engineering contribution with a plausible adaptive MPM scheme, but the core resampling details and one-scene validation need to be documented before I'd trust it for long-horizon tasks.","tokens_in":14744,"tokens_out":1942,"would_cite":false,"duration_ms":20653,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"UBSoft simulates unbounded soft environments for robot learning by keeping fine material resolution near the robot and coarse resolution everywhere else, cutting storage and compute by over an order of magnitude while preserving dynamics…","keywords":["soft-body simulation","material point method","spatially adaptive resolution","robot skill learning","trajectory optimization","sim-to-real transfer","granular materials","unbounded environments"],"falsifier":"Run a robot walking in a straight line through a large sand scene for many steps while tracking total particle mass and the particle field near the robot; if particles systematically vanish or accumulate as they cross the outer boundary of the moving hierarchy, or if the near-field state diverges from a full-domain high-resolution MLS-MPM reference as the scene grows, the unbounded claim is falsified.","tokens_in":13780,"feed_emoji":"🏜️","tokens_out":7458,"duration_ms":72135,"temperature":0.7,"pith_summary":"This paper argues that robot skill learning in soft environments does not have to be confined to small sandboxes. It introduces a simulation platform in which a Material Point Method grid is nested around the robot, with fine cells and small particles near the agent and coarser cells and larger particles farther away; as the robot moves, the hierarchy moves and particles are split or merged to match. The paper claims this spatially adaptive scheme cuts storage and computation by more than an order of magnitude on large soft scenes, and in a falling-cube accuracy test produces dynamics closer to a 256-resolution reference (Chamfer distance 5.32, a point-cloud similarity measure) than a uniform 128-resolution simulation does (12.97) while running in less time. It also reports a benchmark of eight manipulation and locomotion tasks in sand and snow, where sampling-based trajectory optimization generally outperforms reinforcement learning and differentiable-physics gradients, and shows that trajectories optimized in simulation transfer open-loop to a real robot arm for sand painting and scooping.","feed_headline":"Robot-centered grid makes soft-scene simulation 10x cheaper","feed_subtitle":"UBSoft keeps fine detail near the robot, coarse material far away, and transfers learned trajectories to a real arm.","key_machinery":"The carrying mechanism is a moving hierarchical grid paired with a particle split-and-merge rule. Nested cubic grids with side lengths doubling from a base cell size $l_0$ are centered on the active robot; finer grids and smaller particles are used near the robot, coarser grids and larger particles farther out. During each step, particles that cross into a finer layer are split into smaller particles with momentum conserved, and particles in a coarse layer are merged by centroid position and averaged state, so the average particle count per cell stays roughly constant. This hierarchy carries the argument because it lets the platform spend compute and memory only where the agent currently is, while still retaining a coarse record of the rest of the environment and of past interactions.","core_discovery":"The paper's central claim is that an unbounded soft environment can be simulated efficiently by concentrating resolution where it matters. The authors build the platform on the Moving Least Squares Material Point Method, with Lagrangian particles carrying position, velocity, deformation gradient, and volume, and Eulerian grids handling interaction. Their spatially adaptive scheme places nested grids of side length $l_i=2^i l_0$ around the robot, so the finest grid sits at the agent and coarser grids extend outward. Particles in coarse regions are represented by larger particles; when the moving hierarchy overtakes them they are split into smaller particles with momentum conserved, and when small particles accumulate in a coarse cell they are merged into larger ones with averaged state. The paper reports that this reduces storage and time by over an order of magnitude relative to uniform MLS-MPM as scene scale grows, while remaining closer to the fine-resolution reference than a uniform medium-resolution run. The same engine is fully differentiable, supports rigid-body coupling, and the paper demonstrates that one optimized trajectory for sand painting and one for scooping transfer directly to a real robot arm.","pith_inferences":["The same agent-centered resolution idea could carry to particle-based fluids, debris, or coupled rigid-soft scenes where remote material still flows slowly toward the robot; the paper does not test those settings.","The paper argues against freezing distant particles, but does not directly compare against a freeze heuristic; a head-to-head would isolate how much accuracy comes from letting far material evolve slowly at coarse resolution.","Because observations are formed by stratified sampling from hierarchical layers with a fixed total particle budget, the platform could in principle be extended to arbitrarily large scenes with no growth in observation dimension; the paper does not claim this explicitly."],"forward_implications":["Soft-material scenes that would require billions of particles can be simulated faster than real time on current laptop GPUs, opening long-horizon training for locomotion across sand dunes or snowfields.","Memory use stops scaling with total scene volume and instead scales with the resolution layers around the agent, so expanding the simulated world adds mostly coarse particles.","For obtaining a single successful trajectory, sampling-based trajectory optimization (CMA-ES) is the stronger default choice across most of the benchmark tasks.","Differentiable physics gradients remain useful, but mainly for contact-rich, subtly shaped tasks such as sand painting; long-horizon backpropagation through clipping degrades them elsewhere.","Open-loop trajectories optimized in simulation can execute on a real robotic arm for sand writing and scooping, which is evidence of a small sim-to-real gap."],"supporting_citations":[{"why":"Supplies the MLS-MPM method that the spatially adaptive scheme modifies and the uniform-resolution baseline it is compared against.","marker":"[9]"},{"why":"Provides the differentiable, GPU-parallel programming environment in which the engine is implemented.","marker":"[19]"},{"why":"Supplies the Drucker-Prager plasticity model used for sand in the benchmark and sim-to-real experiments.","marker":"[53]"},{"why":"Supplies the elastoplastic snow model used in the locomotion tasks.","marker":"[54]"},{"why":"Represents the height-field approach to large sand scenes that the paper argues cannot capture true 3D soft-material behavior.","marker":"[5]"},{"why":"Represents prior adaptive discrete-continuum coupling for granular media that the paper says still leaves simulation costly.","marker":"[8]"},{"why":"The sampling-based trajectory optimizer that achieves the best rewards on most manipulation and locomotion tasks.","marker":"[55]"},{"why":"One of the model-free reinforcement learning baselines compared on the task suite.","marker":"[56]"},{"why":"One of the model-free reinforcement learning baselines compared on the task suite.","marker":"[57]"}],"fun_headline_variants":["Adaptive grid makes soft-robot training unbounded and cheap","Robot-focused resolution slashes soft simulation costs","Soft sim scales to unbounded scenes with adaptive grids","UBSoft: unbounded soft scenes for robot skill learning","Simulate huge soft scenes by focusing detail on the robot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands or falls on whether coarse particles far from the robot faithfully represent slowly moving material, and on what happens when particles cross the outermost moving grid; the paper tests the former with only one falling-cube run and never specifies the latter.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive grid makes soft-robot training unbounded and cheap","Robot-focused resolution slashes soft simulation costs","Soft sim scales to unbounded scenes with adaptive grids","UBSoft: unbounded soft scenes for robot skill learning","Simulate huge soft scenes by focusing detail on the robot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1347,"prompt_tokens":1015,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":254}},"tokens_in":631,"tokens_out":332,"duration_ms":4074,"temperature":1.0,"reasoning_tokens":254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:12:41.941556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a robot walking in a straight line through a large sand scene for many steps while tracking total particle mass and the particle field near the robot; if particles systematically vanish or accumulate as they cross the outer boundary of the moving hierarchy, or if the near-field state diverges from a full-domain high-resolution MLS-MPM reference as the scene grows, the unbounded claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MLS-MPM method that the spatially adaptive scheme modifies and the uniform-resolution baseline it is compared against."},{"cited_title":"Stomakhin, C","cited_arxiv_id":null,"evidence_quote":"Supplies the elastoplastic snow model used in the locomotion tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the height-field approach to large sand scenes that the paper argues cannot capture true 3D soft-material behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents prior adaptive discrete-continuum coupling for granular media that the paper says still leaves simulation costly."},{"cited_title":"Hansen and A","cited_arxiv_id":null,"evidence_quote":"The sampling-based trajectory optimizer that achieves the best rewards on most manipulation and locomotion tasks."}],"review_version":1}