{"id":"78bb249a-763d-4768-b1db-1c5cbe66d27d","arxiv_id":"2506.03896","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"FLIP calibrates powder simulations using angle-of-repose measurements and trains a curriculum RL policy; on a real robot it achieves 2.12±1.53 mg dispensing error versus 6.11±3.92 mg for domain randomisation.","lead":"FLIP calibrates robot training simulations using the angle of repose of real powders, then trains a reinforcement learning policy with a curriculum that gradually introduces stickier powders. On a real robot, the curriculum-trained policy dispenses powders with about one third the error of a domain-randomization baseline, including for powders more cohesive than any seen in training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AoR-only calibration leaves dynamic dispensing behavior unconstrained; the paper's central sim-to-real claim is not backed by dynamic validation.","rationale":"Agree with the reader's conditional verdict. The paper is a solid empirical study with several independent strengths: the automated AoR system is validated against manual measurements (Table II), Bayesian optimisation retains multiple acceptable parameter sets, and the real-robot evaluation includes in-distribution and out-of-distribution powders as well as an untrained target mass. The comparison with reverse curriculum is a useful control showing that ordering matters. However, the strongest claim is about generalisation enabled by flowability-informed simulation, and the evidence for the simulation's fidelity is limited to the static angle of repose. Section III-A.2 itself disclaims physical accuracy of the calibrated micro-parameters, and Section III-B.2 frames the transfer as a hypothesis. Because the dispensing task is dynamic, matching one quasi-static aggregate observable leaves the dynamic response largely unconstrained; this is a genuine correctness risk, not merely a disagreement with conventions. The proposed dynamic-fidelity test directly probes whether the simulator's calibrated parameters predict real mass release during the learned actions. The secondary concerns already noted by the reader, best-policy selection and the absence of significance tests, reinforce the conditional verdict but are not the decisive issue. No change to the reader's CONDITIONAL verdict is required.","tokens_in":11392,"tokens_out":6896,"duration_ms":72083,"concrete_test":"For each real powder used in training (at least sand, sugar, salt), take the calibrated parameter sets in Theta_F from Section III-B.2 and execute a fixed scripted action sequence (identical shake amplitudes and incline angles) both in simulation and on the physical robot, recording cumulative dispensed mass versus step. Compare the real and simulated mass curves; if the mean absolute difference in intermediate masses exceeds roughly 2-3 mg or the distributions do not overlap, static AoR matching has not transferred the dynamics required for weighing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the AoR-calibrated simulator reproduces the dynamic powder behavior the policy actually controls: the mass released per shake/incline when vibrating and tilting the spoon. Section III-A.1 and Algorithm 1 optimise simulation parameters so that the simulated static AoR matches the real value, and Section IV-B validates only the final pile geometry (Table III). The weighing task, however, is dynamic: dispensing depends on transient jamming, avalanching, cohesion, and adhesion under shaking and inclination. Section III-A.2 explicitly concedes the simplifying assumptions (spherical uniform particles, PBD) and states that optimal parameters 'may not directly correspond to their real-world physical counterparts' because the objective is only to reproduce macroscopic pile behavior. Since flowability is emergent and AoR is a single quasi-static scalar, many parameter sets can produce the same pile angle while having very different dynamic responses (parameter degeneracy). The paper reports no direct comparison of any dynamic flow property, such as mass-flow curves, discharge rate under vibration, or avalanche behavior, between simulation and reality. Consequently, the improved real-world dispensing of FLIP(Curriculum) could be attributable to the curriculum ordering or to robustness induced by training, rather than to dynamic fidelity of the AoR-calibrated simulator. This is the load-bearing weak point in the claim that AoR calibration yields material-specific training environments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FLIP, a framework for robotic powder weighing that uses the angle of repose (AoR) as a flowability metric to calibrate physics-based powder simulations via Bayesian optimization. The calibrated simulators are then used to train a reinforcement learning policy for the weighing task, with a curriculum that progressively increases material cohesiveness. The policy is transferred zero-shot to a real Franka robot and evaluated on five powders, including two with AoR values outside the training range. The authors report that the curriculum-based FLIP achieves a mean dispensing error of 2.12±1.53 mg, outperforming a domain-randomized baseline (6.11±3.92 mg) and other ablations.","tokens_in":11677,"tokens_out":5775,"duration_ms":54219,"significance":"If the central claim is upheld, the work is a meaningful contribution to sim-to-real transfer for granular material manipulation, combining a physically interpretable material descriptor (AoR) with simulation calibration and curriculum learning. The automated AoR measurement is validated against manual measurements, and the simulation calibration achieves low pile-geometry errors for the five tested powders. The paper also provides a useful comparison of curriculum orderings and a domain-randomization baseline. However, the significance is currently tempered by the lack of dynamic validation of the calibrated simulator and by evaluation choices that weaken the statistical strength of the real-world results.","major_comments":[{"comment":"The simulation calibration and validation focus exclusively on the static angle of repose and final pile geometry (Table III), but the dispensing task itself is dynamic: the policy controls shaking and inclination to transfer a target mass, and the outcome depends on transient jamming, avalanching, and cohesion under motion. The paper presents no direct comparison of any dynamic flow property (e.g., discharged mass over time, discharge rate under the shake/incline protocol, or avalanche behavior) between simulation and reality. Since AoR is a single quasi-static scalar and many parameter combinations can reproduce it, the central claim that the AoR-calibrated simulator generates 'accurate training data' for the weighing task is not established. The authors should either measure a dynamic flow property in both domains or substantially temper the claim that AoR calibration is responsible for the sim-to-real transfer.","section":"Section III-A.2 and IV-B"},{"comment":"The real-world evaluation protocol selects the best-performing simulation policy out of five seeds for each method before zero-shot transfer, and then each condition is evaluated with only five real trials. This double selection can bias the reported means, and no statistical significance test or confidence interval is provided for the headline comparison (FLIP(Curriculum) 2.12±1.53 mg vs Domain Randomisation 6.11±3.92 mg). The empirical claim would be much stronger if results were reported for all seeds (or a fixed selection rule were used) and if between-method differences were tested for significance.","section":"Section IV-D"},{"comment":"The formula AoR = 2h/d_base is the tangent of the angle of repose, not the angle itself; the values in Table II are reported in degrees (e.g., 28.55°), which is inconsistent with the formula as written. If the implementation used arctan(2h/d_base), this should be stated explicitly; otherwise the calibration target is mis-specified and the numeric AoR values in the paper are not reproducible.","section":"Equation (2), Section III-A.1"}],"minor_comments":[{"comment":"The acceptance threshold is stated as T_AoR = 1.5° in Section IV-B, but Algorithm 1 line 13 accepts a parameter set when E < 1; please reconcile these values.","section":"Algorithm 1 and Section IV-B"},{"comment":"The paper states that the number of optimised data points per flowability level is set to 7, but it does not explain how these 7 parameter sets are chosen from the top 10 retained by the Bayesian optimization; please clarify the selection rule.","section":"Section IV-C"},{"comment":"There are several typographical errors, including 'signficantly' in the Introduction and 'maneouver' in Section IV-A.2; a careful proofread is recommended.","section":"Throughout"},{"comment":"The phrase 'the first successful application of flowability-informed reinforcement learning' is too strong; consider softening to 'a demonstration' in light of the limited evidence and the issues raised above.","section":"Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a novel idea with promising preliminary results, but the central sim-to-real claim rests on the assumption that AoR calibration transfers to dynamic dispensing behavior, which is not directly validated. The evaluation also has a selection-bias issue that weakens the quantitative comparison. These concerns are addressable with additional experiments (dynamic flow comparisons, all-seed reporting, significance testing), so I would encourage a major revision rather than rejection. The paper is within scope for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a solid empirical systems paper that reports a real improvement in robot powder weighing. The new combination is AoR-targeted Bayesian calibration of a particle simulator, followed by a curriculum that trains on progressively less flowable powders. The real-robot numbers are the hook: curriculum FLIP gets 2.12±1.53 mg mean absolute error across all trials, versus 6.11±3.92 mg for domain randomisation, and it generalizes to out-of-distribution cohesive powders and an unseen 20 mg target. That is worth a look if you work on sim-to-real for granular manipulation.\n\nWhat it does well: the automated AoR measurement is validated against manual measurements (average error 0.84°); the BO calibration visibly reduces sim-to-real pile-formation error (Table III); and the ablations with random ordering and reverse curriculum are informative. The fact that the reverse curriculum is the worst policy is a good control — it shows ordering matters, not just the presence of flowability data.\n\nSoft spots, in proportion. The most concrete weaknesses are methodological. First, real-world evaluation uses the best-performing simulation run per method, which inflates the reported advantage and makes the DR baseline look worse than its average run would. Second, five runs per condition, no significance tests; several error bars overlap. Third, and more substantive: the simulation is calibrated to static AoR only, but the weighing task is dynamic — shaking, inclining, jamming, avalanching. There is no direct comparison of simulated versus real mass-flow or discharge behavior, so we cannot tell whether AoR calibration actually buys dynamic fidelity or just a better-conditioned training distribution. The FLIP(Random) versus DR comparison suggests the AoR-informed data helps independent of curriculum, but it doesn't isolate the mechanism. This is a real gap, and the paper's own caveat about non-physical optimal parameters doesn't fully close it. Fourth, no code or data is released, so the numbers are hard to verify.\n\nIs the central claim sound? The empirical claim — curriculum FLIP transfers better zero-shot — holds up reasonably well. The mechanistic claim, that AoR calibration creates material-specific, dynamically faithful environments, is under-supported. That distinction should be made explicit.\n\nBottom line: this deserves a serious referee and probably a solid workshop or conference slot after revisions. The main requests: include significance tests or more runs, report average-over-runs rather than best-policy selection, and add any dynamic validation you can — even simple mass-flow curves would go a long way. I would cite it if I worked in this area.","headline":"AoR-calibrated sims plus a flowability curriculum look genuinely better for robotic powder weighing, but the dynamic-fidelity claim is under-supported and the evaluation has selection and small-N issues.","tokens_in":12215,"tokens_out":2028,"would_cite":true,"duration_ms":22533,"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":"A flowability-calibrated training pipeline cuts real-robot powder weighing error to 2.12 mg, beating a domain-randomized baseline by a factor of three.","keywords":["powder weighing","flowability","angle of repose","sim-to-real transfer","curriculum learning","Bayesian optimization","robotic manipulation","laboratory automation"],"falsifier":"Measure a dynamic flow property, such as the mass dispensed per shake or the funnel discharge rate, for the five powders and for their AoR-matched simulations; if the simulated dynamic values differ from the real ones by more than the run-to-run noise for any powder the policy handles well, the static-AoR calibration premise is falsified.","tokens_in":11187,"feed_emoji":"⚖️","tokens_out":4967,"duration_ms":44895,"temperature":0.7,"pith_summary":"This paper tries to establish that robotic powder weighing can be made reliable across very different powders by putting a measured physical property—flowability, captured by the angle of repose—explicitly into both the simulation and the training schedule. It claims that calibrating a physics simulator's parameters with Bayesian optimization until simulated piles match real powder piles produces training environments from which a policy transfers zero-shot to a physical robot. It further claims that ordering training from free-flowing to cohesive powders yields a 2.12 ± 1.53 mg average dispensing error, well below the 6.11 ± 3.92 mg of a domain-randomized baseline, and that the policy generalizes to cohesive powders and target masses it never saw. If true, flowability is a sufficient macroscopic descriptor to bridge the sim-to-real gap for granular material handling.","feed_headline":"Flowability-calibrated training cuts robot weighing error to 2 mg","feed_subtitle":"A robot trained in angle-of-repose-calibrated simulations beats domain randomization on unseen, sticky powders in zero-shot transfer.","key_machinery":"The central object is the static angle of repose (AoR), the angle of the cone a powder forms when poured through a funnel; it is used as a stand-in for flowability, with higher AoR meaning less flowable powder. Around it FLIP builds a calibration loop: Bayesian optimization searches simulator parameters such as adhesion, cohesion, friction, particle size, and damping to minimize |AoR_real − AoR_sim(θ)|, retaining parameter sets within 1.5° of the real measurement. Those AoR-matched parameter sets define a flowability-conditioned training distribution, and a curriculum mechanism advances training through parameter sets in order of decreasing flowability once the policy meets an error threshold. The machinery carries the argument because it replaces broad parameter randomization with a physically meaningful scalar that groups powders by manipulation difficulty.","core_discovery":"FLIP's central claim is that material flowability, quantified by the static angle of repose (AoR), is the right organizing variable for sim-to-real transfer in powder manipulation. The paper shows that automated AoR measurements (average error 0.84° vs. manual) can drive Bayesian optimization of a position-based-dynamics powder simulator, yielding parameter sets whose simulated static AoR matches five real powders within 0.22–1.37°. Training a soft actor-critic policy on these flowability-conditioned environments, ordered from lowest to highest AoR, produces a controller that dispenses powders with 2.12 ± 1.53 mg mean absolute error on a real robot, compared with 6.11 ± 3.92 mg for domain randomization, and that handles semolina and sodium bicarbonate—materials outside the 28°–37° training range—and a 20 mg target weight never seen in training.","pith_inferences":["The paper leaves implicit that the same Bayesian-calibration loop could be applied to other measurable macroscopic powder properties, such as dynamic discharge rate or compressibility, to build simulators for scooping, pouring, and tableting tasks.","The poor performance of the reverse curriculum hints that the ordering effect may be about learning stable corrective actions on easy powders before facing cohesive ones; a testable extension would vary curriculum thresholds and spacing to isolate that mechanism.","Because only static AoR is calibrated, the framework's guarantees are bounded by that descriptor; a natural stress test is to shift powder flowability by humidity and check whether re-measuring AoR and re-calibrating restores policy accuracy."],"forward_implications":["A single macroscopic measurement, taken automatically in minutes, can calibrate a powder simulator well enough for zero-shot policy transfer.","Ordering training data by flowability matters: the curriculum beats both random and reverse-curriculum ordering, so the easy-to-hard sequence is itself a source of robustness.","The trained policy generalizes to out-of-distribution cohesive powders and to an untrained target mass, implying the learned controller is not tied to specific materials or setpoints.","All policy training happens in simulation, and the real robot inherits the policy without fine-tuning, making the pipeline practical for laboratory deployment."],"supporting_citations":[{"why":"Defines the powder-weighing task formulation, the shake-and-incline action space, the SAC training setup, and the domain-randomized baseline that FLIP is compared against.","marker":"[6]"},{"why":"Supplies the curriculum-learning strategy used to order flowability levels from easy to hard.","marker":"[8]"},{"why":"Defines domain randomization, the parameter-agnostic baseline approach that FLIP is designed to outperform.","marker":"[16]"},{"why":"Shows how granular simulation parameters can be inferred from visual observations, providing the methodological precedent for simulation calibration from real data.","marker":"[18]"},{"why":"Provides the Bayesian optimization routine used to search simulator parameters against the AoR error.","marker":"[21]"},{"why":"Defines the Position-Based Dynamics solver that simulates the powder particle dynamics in the training environment.","marker":"[22]"},{"why":"Provides the task-simulation framework in which the powder-weighing and AoR measurement environments are built.","marker":"[25]"},{"why":"Supplies the Soft Actor-Critic algorithm, the off-policy deep reinforcement learning method used to train the dispensing policy.","marker":"[26]"}],"fun_headline_variants":["Flowability-calibrated sims cut robot powder error to 2 mg","Angle-of-repose data trains robots to weigh sticky powders","Robot powder weighing hits 2 mg error via flowability-aware sims","Sim-to-real powder handling improved by flowability calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole transfer rests on one premise: a powder whose static angle of repose matches the simulation will also flow, shake, and dispense like the real powder during the dynamic weighing motion—but the paper only measures the static pile angle, never a dynamic flow property.","fun_headline_variants_meta":{"raw":{"variants":["Flowability-calibrated sims cut robot powder error to 2 mg","Angle-of-repose data trains robots to weigh sticky powders","Robot powder weighing hits 2 mg error via flowability-aware sims","Sim-to-real powder handling improved by flowability calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1963,"prompt_tokens":950,"completion_tokens":1013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":940}},"tokens_in":566,"tokens_out":1013,"duration_ms":9333,"temperature":1.0,"reasoning_tokens":940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:53:36.943181+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a dynamic flow property, such as the mass dispensed per shake or the funnel discharge rate, for the five powders and for their AoR-matched simulations; if the simulated dynamic values differ from the real ones by more than the run-to-run noise for any powder the policy handles well, the static-AoR calibration premise is falsified.","supporting_citations":[{"cited_title":"Learning robotic pow- der weighing from simulation for laboratory automation,","cited_arxiv_id":null,"evidence_quote":"Defines the powder-weighing task formulation, the shake-and-incline action space, the SAC training setup, and the domain-randomized baseline that FLIP is compared against."},{"cited_title":"Inferring the material properties of granular media for robotic tasks,","cited_arxiv_id":null,"evidence_quote":"Shows how granular simulation parameters can be inferred from visual observations, providing the methodological precedent for simulation calibration from real data."},{"cited_title":"Practical bayesian opti- mization of machine learning algorithms,","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian optimization routine used to search simulator parameters against the AoR error."},{"cited_title":"Orbit: A unified simulation framework for interactive robot learning environments,","cited_arxiv_id":null,"evidence_quote":"Provides the task-simulation framework in which the powder-weighing and AoR measurement environments are built."},{"cited_title":"Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,","cited_arxiv_id":null,"evidence_quote":"Supplies the Soft Actor-Critic algorithm, the off-policy deep reinforcement learning method used to train the dispensing policy."}],"review_version":1}