{"id":"ce7cd858-cd03-4484-93cd-7d3d79fdf277","arxiv_id":"2607.20912","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A robot manipulation policy that predicts an impedance-admittance switch ratio alongside virtual targets and stiffness achieves higher success and fewer contact failures in rigid-contact tasks.","lead":"Researchers trained a robot policy that predicts not only where to move and how stiff to be, but also which low-level control mode to use, switching between precise tracking and stable contact. This reduced tool breakage and force oscillations in rigid-contact tasks like flipping a stiff box and pressing along a line.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reactive switch-ratio baselines and the label-force confound leave the 'predicting ahead' claim unsupported; Eq. 6 cannot produce the reported pre-contact n≈0 behavior.","rationale":"The reader correctly identified the stability of time-varying switching as a genuine gap, but the more load-bearing issue for the paper's central claim is experimental identification: the switch-ratio labels are reactive functions of force, yet the paper claims predictive, ahead-of-contact behavior. The reported pre-contact n≈0 contradicts the label definition unless a confound (human guidance force in teleoperation) or an undisclosed temporal processing step is present. Without a reactive force-threshold baseline, the benefit of learned prediction is not established. This concern does not overturn the reader's CONDITIONAL verdict; it strengthens the need for additional experiments and documentation, so the verdict remains CONDITIONAL. The stability concern is secondary but also worth addressing, as the implementation as duty-cycle switching may not be covered by Ott et al. [29].","tokens_in":11460,"tokens_out":7115,"duration_ms":68102,"concrete_test":"Run 20 evaluation trials of a reactive baseline that computes n_t online from the current measured |F_ext| using Eq. 6 with the same f_min/f_max used for URF labels, combined with the same predicted virtual target and K_t (or URF's stiffness predictions) under the unified controller. Compare success rate, PFGR, ORMS, and CFR with URF. Additionally, inspect the demonstration force traces and label pipeline to verify whether pre-contact labels are n≈1 or n≈0, and report the exact f_min/f_max values used in both tasks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is a policy that predicts the switch ratio n_t, and the key evidence is that URF lowers n before contact, avoiding force buildup. But the supervision signal for n_t (Eq. 6) is a function of the instantaneous measured force magnitude |F_ext|: low force yields n=1, high force yields n=0. In the approach phase of a demonstration, before contact, |F_ext| is near zero, so the label is n≈1. A diffusion policy trained with the loss in Eq. 7 to regress these labels should therefore output n≈1 during approach. Yet Section III-B-1 reports that 'Before contact, the policy has already set n≈0' (Fig. 3c). This is only possible if the force used to build labels includes human guidance force during teleoperation (so |F_ext| is large during the human's pushing approach), or if the labels are temporally shifted or processed in a way not described. Neither is stated. The paper also omits the f_min/f_max values and never compares against a reactive force-threshold policy that sets n_t from current force using Eq. 6. Without such a baseline, the observed success rate cannot be attributed to 'predicting' controller behavior before force buildup; a reactive mapping from force to n could reproduce the results. The central claim in the abstract therefore remains underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes URF, a framework in which a diffusion policy predicts not only a virtual target and stiffness matrix but also a scalar impedance-admittance switch ratio n_t. The low-level controller switches between admittance and impedance behaviors according to n_t, with labels for n_t derived from measured force magnitude via Eq. (6). The method is evaluated on two rigid-contact tasks, box-flipping and line-pressing, with 20 trials per condition and compared against a standard diffusion policy, the Adaptive Compliance Policy (ACP), and two fixed-switch ablations. Reported results show URF achieving 90% success in box-flipping versus 25% for ACP, and 100% success in line-pressing versus 0% for ACP, with lower peak force growth rates and lower force oscillations. The central claim is that predicting the controller mode in addition to compliant actions improves stability and task success in rigid contact.","tokens_in":11747,"tokens_out":4980,"duration_ms":49678,"significance":"If the claims are supported, the paper makes a useful contribution: it couples learned action prediction with the low-level control mode, which is often treated as a fixed component, and it proposes a practical force-based labeling scheme that avoids explicit environment-stiffness estimation. The rigid-contact task design is appropriate for exposing controller-related failure modes, and the fixed-n ablations help show that adaptive switching matters. The authors are also transparent about two limitations in Section IV. However, the key novelty—that the policy predicts switch ratios before contact and thereby avoids force buildup—is currently under-supported because the supervision signal is a deterministic function of current force, no reactive force-threshold baseline is tested, and the relevant bounds and statistical uncertainty are not reported. The stability of time-varying switching is asserted by citation rather than analyzed or empirically demonstrated.","major_comments":[{"comment":"Equation (6) assigns n_t=1 when |F_ext| is below f_min, so approach-phase samples before physical contact should have labels near 1. Yet Section III-B-1 reports that 'before contact, the policy has already set n≈0' (Fig. 3c). The manuscript does not explain how pre-contact n≈0 labels arise. If the measured force used for labeling includes the human guidance force during direct teaching, then |F_ext| is large during the approach and the label is not a pure contact-loading signal; if the labels are temporally shifted, filtered, or otherwise processed, that processing is not described. Because the abstract's 'predicting ahead' claim depends on pre-contact reduction of n, this inconsistency needs to be resolved explicitly.","section":"II-B-3 and Fig. 3(c)"},{"comment":"No baseline uses the force-to-switch mapping of Eq. (6) as a reactive controller. Since the supervision signal is exactly a threshold on current force magnitude, a reactive policy that sets n_t from the current |F_ext| is a natural control condition and would test whether predicting switch ratios ahead of time adds value beyond the current force observation. Without such a baseline, the observed improvements in PFGR, CFR, and safety stops cannot be attributed to 'predicting' controller behavior before force buildup; a reactive force-threshold rule could reproduce the same qualitative behavior.","section":"III-B and Tables I-II"},{"comment":"The stability claim 'has been shown to yield stable switching behavior' cites Ott et al. [29], but the cited analysis is not extended to the setting used here, where K_t and n_t are time-varying, policy-generated commands updated at low rate (policy inference) while the low-level controller runs at 1,000 Hz. No passivity or Lyapunov argument is provided for this time-varying switching, and the experiments use only 20 trials per condition. The manuscript should either extend the stability argument to the actual control architecture or provide empirical evidence, such as boundedness of force and velocity across trials and seeds, that time-varying switching is stable.","section":"II-A"},{"comment":"The values of f_min and f_max are never reported, and the controller gains M_d, D_d, K_p, and K_v are also not given. Since Eq. (6) is the core supervision signal for the switch ratio and the low-level unified controller is the main contribution, these missing values prevent reproducibility and make it impossible to interpret the force-threshold behavior that the paper claims to learn. Please report these parameters for both tasks.","section":"III-A and Eq. (6)"},{"comment":"All success rates are computed over 20 trials with no confidence intervals, no information about random seeds, and no per-trial breakdown. With n=20, the difference between 90% and 60% success has a wide binomial confidence interval, and the reported PFGR and ORMS means and standard deviations do not account for seed variability. The claim that URF 'achieves higher task success rates' would be substantially strengthened by binomial confidence intervals, multiple seeds, and per-trial data for the safety-related metrics.","section":"Tables I-II"}],"minor_comments":[{"comment":"The sentence 'The three URF variants have similar PFGR values, which suggesting that impedance-like behavior near contact is sufficient' contains a grammatical error; 'which suggesting' should be 'which suggests' or 'suggesting'.","section":"III-B-1"},{"comment":"The manuscript states that force/torque measurements are low-pass filtered and downsampled to 100 Hz for data collection and policy input, but the cutoff frequency of the low-pass filter is not reported. Please specify it for reproducibility.","section":"III-A"},{"comment":"The force-norm traces in Figs. 3b and 4b appear to come from single representative trials; the captions should state this explicitly and indicate how the shown trial was selected, since the tables already report aggregate statistics.","section":"Fig. 3 and Fig. 4"},{"comment":"The observation encoding is described as using two RGB frames and three end-effector poses, but the time horizon and sampling rate of these observations are not stated; please clarify whether these are consecutive timesteps at 100 Hz or spread over a longer window.","section":"II-B-1"},{"comment":"The project page is mentioned in the abstract but not included in the reference list; a footnote or reference with the URL would help readers access supplementary videos and code.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is interesting, but the evidence for the central 'predicting ahead' claim is currently confounded by the force-derived labels and the absence of a reactive force-threshold baseline. If the authors can add that baseline, report f_min/f_max and controller gains, provide confidence intervals, and clarify the labeling pipeline, the contribution would be much stronger. I would not recommend rejection at this stage, but the revision needs to address the confounding directly rather than only adding more trials."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"URF is a well-executed systems paper with a genuinely new action dimension: predicting an impedance-admittance switch ratio alongside virtual targets and stiffness. The real-robot experiments on box-flipping and line-pressing are clean, the ablations are informative, and the writing is honest about limitations. But the central claim that the policy 'predicts ahead' rather than reacting to force is not supported by the evidence as presented.\n\nThe strength is the idea itself and the force-based labeling trick. The comparison with ACP and fixed-switch ablations shows clear practical gains, and the force traces explain the failure modes well. The limitations section admits the f_min/f_max sensitivity and the axis-shared switch ratio.\n\nThe soft spots are substantial. Eq. 6 defines the switch-ratio label as a deterministic function of instantaneous measured force magnitude. The paper reports that before contact the policy already sets n≈0, which would require a non-zero measured force during approach. The most plausible explanation is that the force sensor also reads the human's teaching force, but the paper never says this. If the labels are effectively a reactive threshold on force, the 'predicting ahead' advantage needs a baseline that sets n_t from current force via Eq. 6. Without it, the observed success rates could come from a reactive mapping, and the central contribution is underdetermined. The paper also omits the f_min/f_max values, gives only 20 trials per method with no confidence intervals, and releases no code or data. The stability of the time-varying switch ratio is asserted by citing Ott et al., but that analysis does not cover low-rate policy commands; the experiments show it works on these two tasks, but there is no formal or empirical analysis beyond that.\n\nThe stress-test note is on point: the label-force confound is real and needs to be addressed directly.\n\nMy take: this is a conditional accept at best. The core idea deserves serious peer review, but the missing reactive baseline and the label-detail omissions prevent the central claim from being verified. I'd send it to review, expecting major revision. If the authors add a reactive force-threshold baseline, report f_min/f_max, and release code/data, the paper could be a solid contribution.","headline":"A promising controller-mode prediction idea whose central 'predicting ahead' claim is undercut by a label-force confound and a missing reactive baseline.","tokens_in":12289,"tokens_out":3204,"would_cite":false,"duration_ms":27789,"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":"URF claims that contact-aware manipulation policies should also predict the low-level controller mode, and shows that force-labeled impedance-admittance switching raises task success and cuts contact-induced failures.","keywords":["unified impedance-admittance control","contact-aware manipulation","diffusion policy","switch ratio","force-supervised labeling","virtual target","stiffness learning","rigid contact"],"falsifier":"Run URF on the line-pressing task twice with the same predicted virtual-target and stiffness sequences, once with the learned switch ratios and once with each step's average switch ratio substituted. If the averaged version keeps 100% success and low force oscillation, the paper's claim that predicting the switch ratio carries the result is unsupported; if it fails or oscillates, adaptive switching is the load-bearing component.","tokens_in":11288,"feed_emoji":"🤖","tokens_out":5988,"duration_ms":51927,"temperature":0.7,"pith_summary":"URF claims that in rigid-contact manipulation, a policy should predict not only the compliant action, namely a virtual target and stiffness, but also the low-level controller mode used to execute it. To make this learnable, it derives an impedance-admittance switch-ratio label from the measured force magnitude, replacing the unavailable ground-truth environment stiffness with a conservative force-based rule. On box-flipping and line-pressing, the framework posts higher success rates than an admittance-only compliance policy while avoiding its characteristic failures: rapid force buildup, large force oscillations, tool breakage, and robot safety stops. The message is that action prediction and interaction control should be designed as one coupled system.","feed_headline":"Robot policy that switches controller mode wins 90% of box flips","feed_subtitle":"URF predicts impedance-admittance switching from force history, cutting tool breakage and safety stops.","key_machinery":"The load-bearing object is the unified impedance-admittance controller [29], repurposed to accept time-varying inputs: the virtual target $x^v_t$, the stiffness matrix $K_t$, and the switch ratio $n_t$. Within each control period, $n_t$ sets the fraction of time spent in the admittance phase, with state mapping at the switch to keep commanded force and its derivative continuous. The other pillar is the force-based label in Eq. (6): $n_t$ is a piecewise-linear function of the measured force norm, staying near $1$ below $f_{\\min}$ and dropping to $0$ above $f_{\\max}$. A conditional diffusion policy with a U-Net denoiser generates 16-step action chunks of $(x^v_t, K_t, n_t)$ from image, pose, and force-history observations in a receding-horizon loop.","core_discovery":"The paper's central claim is that contact-aware policies should output the controller behavior alongside the compliant action, and that this is practically trainable. Specifically, URF predicts the triple ($x^v_t$, $K_t$, $n_t$), a virtual target pose, a stiffness matrix, and a scalar impedance-admittance switch ratio, where $n_t=1$ is pure admittance for accurate tracking and $n_t=0$ is pure impedance for stable stiff contact. Because demonstrations lack environment-stiffness annotations, the authors label $n_t$ from the measured force norm via a piecewise-linear rule between user-set bounds $f_{\\min}$ and $f_{\\max}$. In box-flipping, URF reaches 90% success versus 25% for the admittance-only adaptive compliance policy; in line-pressing it sustains contact in every trial while the admittance baseline always ends in a safety stop. The paper therefore argues that stable contact alone is insufficient, because accurate approach and tangential tracking require the controller mode to adapt during execution.","pith_inferences":["An unstated extension is that the same force-based labeling could be applied to other variable-compliance controllers, such as variable admittance gains, turning controller-aware policy learning into a general recipe rather than a single-controller trick.","If the learned switch ratio tracks environment stiffness, then $n_t$ is an implicit stiffness estimator, which could be tested in simulation where ground-truth environment stiffness is known by comparing predicted $n_t$ with the optimal mode.","The authors note that the switch ratio is shared across all control axes; a natural per-axis extension would let one axis stay stiff while another stays compliant, which is exactly what insertion tasks require.","The manual force bounds $f_{\\min}$ and $f_{\\max}$ are a likely bottleneck, and the paper itself proposes letting richer scene understanding infer them, a testable next step that would remove the main user-defined parameter."],"forward_implications":["If URF is right, action spaces for contact-rich imitation should include the controller mode, because the same virtual target and stiffness produce different physical outcomes under different low-level controllers.","Force-magnitude labeling offers a route to controller-aware supervision without estimating environment stiffness, which is usually unobservable from demonstrations.","The fixed-switch ablations imply that neither pure impedance nor a fixed 50/50 blend is enough: tasks that need accurate placement before contact and stable force after contact require the switch ratio to adapt over time.","Because the switch ratio changes before large forces appear, predicted from images and force history, the policy is not merely a force threshold but anticipates approaching contact conditions.","The two demonstrated tasks suggest the approach transfers across short impulsive contact, box flipping, and sustained contact, line pressing."],"supporting_citations":[{"why":"Supplies the unified impedance-admittance controller whose smooth mode switching URF adopts as its low-level execution layer.","marker":"[29]"},{"why":"Supplies the adaptive compliance labeling procedure from which URF derives virtual-target and stiffness labels for demonstrations.","marker":"[25]"},{"why":"Supplies the diffusion-policy action-chunk generation architecture that URF extends to controller-aware action triples.","marker":"[5]"},{"why":"Supplies the denoising diffusion noise-prediction objective used to train the action-chunk policy.","marker":"[11]"}],"fun_headline_variants":["URF unifies impedance and admittance, hits 90% flips","Robot policy switches controller mode based on force","One policy predicts target, stiffness, and mode switch","Admittance for tracking, impedance for contact: URF combines","From safety stops to full trials: URF's mode switch works"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The controller stays stable when the stiffness matrix and switch ratio are time-varying commands generated by a policy running at 100 Hz while the control loop runs at 1 kHz, yet the cited switching-stability result is not extended to this setting.","fun_headline_variants_meta":{"raw":{"variants":["URF unifies impedance and admittance, hits 90% flips","Robot policy switches controller mode based on force","One policy predicts target, stiffness, and mode switch","Admittance for tracking, impedance for contact: URF combines","From safety stops to full trials: URF's mode switch works"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1505,"prompt_tokens":1000,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":616,"tokens_out":505,"duration_ms":4735,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:31:06.626647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run URF on the line-pressing task twice with the same predicted virtual-target and stiffness sequences, once with the learned switch ratios and once with each step's average switch ratio substituted. If the averaged version keeps 100% success and low force oscillation, the paper's claim that predicting the switch ratio carries the result is unsupported; if it fails or oscillates, adaptive switching is the load-bearing component.","supporting_citations":[{"cited_title":"Unified impedance and admittance control,","cited_arxiv_id":null,"evidence_quote":"Supplies the unified impedance-admittance controller whose smooth mode switching URF adopts as its low-level execution layer."},{"cited_title":"Adaptive compliance policy: Learning approximate compliance for diffusion guided control,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive compliance labeling procedure from which URF derives virtual-target and stiffness labels for demonstrations."},{"cited_title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-policy action-chunk generation architecture that URF extends to controller-aware action triples."}],"review_version":2}