{"id":"19073604-37f8-42b2-8c18-54b27910d755","arxiv_id":"2412.13157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A visuotactile estimator with learned uncertainty, trained from privileged-policy data, lets an RL controller push objects successfully under prolonged visual occlusions.","lead":"A robot learns to keep pushing a box to a target even when its own camera view is blocked, using force feedback and a learned estimate of the box's pose and uncertainty. The system, tested in simulation and on a KUKA robot arm, handles prolonged occlusions with a single onboard camera instead of external trackers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim-to-real transfer relies on an unvalidated synthetic occlusion/force model; hardware success rates do not quantify estimator accuracy or uncertainty calibration under real occlusions.","rationale":"The reader's weakest assumption is that the synthetic observation model—freeze-last-pose occlusions, Table 1 noise levels, and the force-noise compensation described in Sec. 6—is representative of the real robot. I agree that this is the most load-bearing point. The paper's central assertion is a sim-to-real transfer claim, and the only direct evidence for it is a set of hardware success counts. Those successes are valuable and should be credited; they are real falsifiable evidence that the system can push under occlusions. However, they do not isolate the mechanism claimed: whether tactile feedback plus learned uncertainty substitutes for visual tracking depends on the estimator being accurate and its uncertainty being calibrated on the physical system. The paper explicitly declines to evaluate uncertainty calibration, and Sec. 6 concedes an unmodeled systematic force signal. An external ground-truth instrument would settle this directly and would also address the stated limitation. Because the reader already issued a CONDITIONAL verdict on substantially these grounds, my stress-test does not move the verdict; it sharpens the condition that would convert it to acceptance: a hardware estimator-calibration study. I do not see an internal inconsistency in Equations (1)-(4) or in the training procedure, and the hardware results prevent me from calling the transfer assumption falsified. The concern is about missing evidence, not demonstrated failure.","tokens_in":12473,"tokens_out":7505,"duration_ms":79703,"concrete_test":"Instrument the physical setup with an external motion-capture system or a second fixed camera that is not used by the policy, and record ground-truth object pose during the natural-occlusion and full-occlusion hardware trials. For each timestep, compute the estimator's translation/rotation error and the calibration of its predicted covariance (e.g., the fraction of true poses inside the predicted 1-sigma ellipse and the negative log-likelihood of the true pose under the predicted Gaussian). Compare these quantities against the simulation test-set results underlying Table 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a policy trained entirely in simulation, on an observation model where occlusions freeze the last detected pose and force noise is additive Gaussian (Table 1), transfers zero-shot to hardware with a simple onboard camera. Section 6 states that the simulator neglects pusher dynamics that produce significant non-contact force readings on the real wrist F/T sensor, and that the authors compensated by adding 'large amounts' of correlated and uncorrelated noise. This is an explicit admission that the force observation model is not a faithful generative model of the hardware sensor. Because the estimator's predicted covariance (Eq. 4) is fed to the policy and used to choose contact strategies (Sec. 5.4), miscalibrated uncertainty could change the policy's behavior in deployment. The positive hardware numbers (19/20, 10/10, 7/10) show the system works in those trials, but without measuring the estimator's pose error or covariance calibration against ground truth on the robot, they do not establish that the simulator's synthetic occlusion and noise model is representative enough for the claimed mechanism to be the cause of success. The Limitations section itself acknowledges that no quantitative evaluation of uncertainty accuracy was performed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a visuotactile state estimation and uncertainty-aware control pipeline for planar non-prehensile manipulation under visual occlusions. The method first trains a privileged RL policy in an occlusion-free simulation, uses its training checkpoints to collect a diverse dataset, and then trains a Bayesian deep learning estimator that outputs both object pose and aleatoric/epistemic uncertainty via MC dropout. The estimator is embedded in the RL loop to train a control policy that receives the estimated pose and covariance as input. In simulation the method achieves 94% success, and with zero-shot sim-to-real transfer it attains 19/20, 10/10, and 7/10 success on a KUKA iiwa with an onboard camera under naturally occurring and human-induced occlusions. The central claim is that explicitly modeling uncertainty, rather than relying on complex external perception, is what makes occlusion-robust behavior possible after transfer.","tokens_in":12705,"tokens_out":4098,"duration_ms":40491,"significance":"If the result holds, the paper makes a valuable contribution: it demonstrates that learned uncertainty from a proprioceptive/tactile estimator can substitute for continuous visual tracking in a contact-rich non-prehensile task, enabling simpler onboard perception setups. The experimental design is generally careful: separate train/validation/test datasets, three RL training seeds, multiple ablations (MSE estimator, likelihood loss, end-to-end RL, behavior cloning, privileged policy with estimator), and hardware trials with randomized initial configurations. The paper also provides reproducible training details in the appendix. The main open question is whether the hardware success is actually attributable to the learned uncertainty mechanism, given that the uncertainty calibration is not evaluated on the robot and the simulation observation model is admittedly approximate.","major_comments":[{"comment":"The hardware section states that the simulator neglects pusher dynamics that cause 'significant force readings outside of contact interactions' on the real F/T sensor and that the authors instead 'relied on adding large amounts of correlated and uncorrelated noise to the force observation.' Section 7 additionally states that 'we omit any quantitative evaluation on the accuracy of the estimated uncertainty.' Because the policy receives the predicted covariance from Eq. (4) as an input (Section 3.3) and the authors attribute the successful behavior to uncertainty-aware contact-face switching (Section 5.4), the hardware results do not currently establish that the learned uncertainty is well-calibrated on the real robot. Without measuring estimator pose error or covariance calibration against ground truth on the hardware, the positive trials only show that the overall system works, not that the synthetic occlusion/noise model is representative enough for the claimed mechanism to be the cause of success. Please add a quantitative hardware evaluation of the estimator's accuracy and uncertainty calibration (e.g., comparing predicted variance with empirical error across trials) or substantively temper the claim that uncertainty modeling is what enables the sim-to-real transfer.","section":"Section 6 and Section 7"},{"comment":"The claim that modeling uncertainty 'significantly improves the state estimator accuracy' is based solely on Table 2, which reports point estimates (Mean L2, RMSE, Mean Abs., RMSE) with no variance, no number of seeds, and no statistical test. Given that the estimator is the foundation of the policy and that the differences between MSE (7.62 mm), Likelihood (6.83 mm), and Ours (4.42 mm) are the basis for this claim, please report standard deviations or confidence intervals across estimator training seeds and test splits, or provide a significance test.","section":"Table 2"},{"comment":"The distinction between πest(ŝ, Σ) and πest(ŝ) supports the paper's emphasis on explicit uncertainty input, but the final difference is only 94% versus 92% and Figure 2 shows overlapping standard deviations across three seeds. The conclusion that 'providing the uncertainty estimate explicitly to the policy leads to slightly improved performance' is not supported by any significance test. Please report per-seed final success rates and perform a simple statistical comparison (e.g., bootstrap or a paired test across seeds), or soften the claim to a trend.","section":"Section 5.2 and Figure 2"},{"comment":"The synthetic occlusion model freezes the last observed object pose for a duration sampled from N(10, 5^2) s and adds a binary indicator ξ. On the real hardware, occlusions arise when AprilTag markers leave the camera field of view or are obstructed by the pusher. The paper does not specify how the real detection pipeline behaves during occlusion (does it output the last detected pose, a missing flag, or a zero vector?) and whether that behavior matches the synthetic 'frozen pose' model. This is a potential sim-to-real gap in the observation model that affects the estimator's inputs and the policy's behavior. Please clarify the actual detection pipeline and, ideally, report the empirical distribution of detection outages on the hardware to justify the synthetic model.","section":"Section 3.2 Data Processing and Section 6"}],"minor_comments":[{"comment":"Table 1 lists the occlusion duration as N(10, 5^2) s, but Section 5.3 fixes the duration for evaluation scenarios. Please clarify whether the table refers only to training, and separately state the distribution of the correlated noise component referenced in Section 4.","section":"Table 1"},{"comment":"Equation (2) gives the sample covariance without Bessel's correction (the outer factor is 1/M rather than 1/(M−1)). For M=50 this is numerically negligible, but the formula should be flagged as the biased estimator or corrected for formal consistency.","section":"Eq. (2)"},{"comment":"The paper says '300 uniformly spaced privileged policy checkpoints' are used for data collection, but it is not stated how spacing is measured (e.g., every N optimizer steps) or whether the checkpoints include the final trained policy. Please specify the selection criterion.","section":"Section 4 (State Estimator)"},{"comment":"The hardware success counts 19/20, 10/10, and 7/10 are small samples; please report binomial (e.g., Clopper-Pearson) confidence intervals to convey the precision of these rates.","section":"Section 6"},{"comment":"The comparison with prior model-based estimators [36, 37] is qualitative ('similar or better accuracy'); please provide the relevant numeric accuracies from those references so the reader can verify the claim.","section":"Section 5.1"},{"comment":"In the sentence 'We find that πpriv(st) makes 2.68±1.53, while πest(ŝt, Σt) makes 4.22±2.67 contact switches per episode,' the notation mixes πpriv(st) and πpriv(ŝt); please use consistent notation for the privileged policy with and without the estimator.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's positive hardware results and extensive ablations make it a solid candidate, but the central mechanism claim (that explicit uncertainty modeling drives the occlusion robustness) is not directly validated on hardware, and the estimator comparison in Table 2 lacks variance information. These are fixable with additional experiments or by tempering the claims, so I recommend major revision rather than rejection. The fit to the journal's scope is good, assuming the venue values sim-to-real transfer in contact-rich manipulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: it trains a visuotactile state estimator with Bayesian uncertainty and an uncertainty-aware RL policy, and shows the combination survives prolonged occlusions in planar pushing, both in simulation and on a real KUKA with an onboard camera. The building blocks are all established, but the explicit treatment of occlusions as a first-class problem in non-prehensile manipulation, and the evidence that uncertainty input changes the learned contact strategy, is new. I would not desk-reject this.\n\nWhat is genuinely good: the privileged-policy checkpoint idea for collecting diverse interaction data is simple and effective; the ablations in Table 2 and Figure 2 cleanly attribute gains to aleatoric and epistemic uncertainty; and the hardware numbers (19/20, 10/10, 7/10) are solid for zero-shot sim-to-real. The comparison against end-to-end RL and behavior cloning is fair and makes the case that explicit state estimation matters here. The limitations section is unusually candid.\n\nThe soft spots are real but not fatal. Table 2 has no error bars or seeds, so the 'significantly improved estimator' claim is under-supported; the simulation evaluation uses the same occlusion distribution the policy was trained on, so it is in-distribution; and, as the paper itself admits, the force observation model deliberately ignores pusher dynamics and compensates with large synthetic noise. The hardware trials show the system works, but they do not measure estimator pose error or covariance calibration on the robot, so the mechanism—that the learned uncertainty is what drives the good behavior—is plausible but not directly confirmed. The policy's observable contact-switching behavior, visible in Figure 4 and the video, is supporting evidence. No code or data are released, which also weakens reproducibility.\n\nThe stress-test note is right that the sim-to-real bridge relies on an unvalidated synthetic occlusion/force model. But the paper does not hide this; it is stated in Section 6 and in the Limitations. Given the positive hardware results, I do not think this is a load-bearing flaw—it is a reason to want more evidence, not a reason to reject.\n\nThis is a paper for robotics researchers working on contact-rich manipulation, occlusion-robust perception, and sim-to-real. I would bring it to a reading group and would cite it. For peer review, yes—send it to referees. It deserves a serious look, with requests for error bars on the estimator table, uncertainty calibration on hardware, and ideally code and data.","headline":"A credible systems paper: the components are known, but the explicit occlusion-focused formulation plus estimator-in-the-loop RL and positive hardware results make it a legitimate new result; the main soft spot is the unvalidated synthetic sensor model in sim-to-real.","tokens_in":13224,"tokens_out":1522,"would_cite":true,"duration_ms":16991,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a robot can keep pushing an object to target under prolonged visual occlusions by learning a visuotactile state estimator that reports its own uncertainty and training the control policy with that uncertainty in the loop.","keywords":["non-prehensile manipulation","planar pushing","visuotactile state estimation","Bayesian deep learning","uncertainty-aware control","reinforcement learning","sim-to-real transfer","visual occlusions"],"falsifier":"Observe whether the estimated uncertainty rises during gradual partial occlusions on hardware; a failure to anticipate full occlusion would indicate the binary freeze-last-pose observation model is load-bearing.","tokens_in":12277,"feed_emoji":"🤖","tokens_out":8237,"duration_ms":68408,"temperature":0.7,"pith_summary":"The paper claims that a robot can keep pushing an object to a target even when its only camera view is occluded, by learning a state estimator that fuses vision, proprioception, and force/tactile readings and by feeding that estimator's uncertainty into the control policy. The estimator is trained on diverse simulated interaction data generated by privileged policy checkpoints, using a Bayesian deep learning loss that outputs both pose and covariance. The policy is then trained with the estimator in the loop, which lets it learn actions that actively reduce uncertainty, such as switching contact faces. On a planar pushing task the method reaches 94% success in simulation and transfers zero-shot to hardware, with 19/20, 10/10, and 7/10 runs under different occlusion scenarios. If this holds, non-prehensile manipulation no longer needs elaborate external tracking rigs, and a simple onboard camera plus force sensing may suffice.","feed_headline":"Uncertainty-aware estimator lets robots push objects they cannot see","feed_subtitle":"Tactile feedback plus learned uncertainty beats end-to-end RL and behavior cloning when vision fails.","key_machinery":"The machinery is a two-stage learning pipeline built on a GPU-accelerated physics simulator. A privileged policy $\\pi_{\\mathrm{priv}}(s_t)$ is trained with PPO in an occlusion-free environment, and equally spaced checkpoints (including untrained and suboptimal ones) are rolled out to collect 750,000 trajectories. Those trajectories are processed with a synthetic observation model that adds correlated and uncorrelated Gaussian noise and occlusions that freeze the last observed pose, with duration sampled from $\\mathcal{N}(10, 5^2)$ seconds. The estimator is an LSTM-based network with MC dropout that predicts the object pose and aleatoric log-covariance; epistemic uncertainty is computed from 50 stochastic forward passes, and total uncertainty is $\\hat{\\Sigma}_t = \\hat{\\Sigma}^{\\mathrm{ale}}_t + \\hat{\\Sigma}^{\\mathrm{epi}}_t$. The control policy, an RL agent with the same recurrent architecture, receives the estimated state and the standard deviations of the total uncertainty as inputs, and is trained by PPO in the occlusion environment with the estimator frozen in the loop.","core_discovery":"The central discovery is that explicitly modeling both aleatoric and epistemic uncertainty in a learned visuotactile state estimator, and providing that uncertainty to a control policy trained with the estimator in the loop, makes planar pushing robust to prolonged visual occlusions. The estimator $f(o_t)$ outputs the object pose mean $\\hat{q}^{\\mathrm{obj}}_t$ and a diagonal covariance $\\hat{\\Sigma}_t$, trained with the negative log-likelihood loss $\\mathcal{L} = \\frac{1}{2}\\ln|\\hat{\\Sigma}^{\\mathrm{ale}}_t| + \\frac{1}{2}(q^{\\mathrm{obj}}_t - \\hat{q}^{\\mathrm{obj}}_t)^\\top(\\hat{\\Sigma}^{\\mathrm{ale}}_t)^{-1}(q^{\\mathrm{obj}}_t - \\hat{q}^{\\mathrm{obj}}_t)$, with epistemic uncertainty added via Monte Carlo dropout. This estimator outperforms an MSE-trained variant (4.42 mm vs 7.62 mm mean translation error) and a vision-only baseline, and the policy $\\pi_{\\mathrm{est}}(\\hat{s}_t, \\hat{\\Sigma}_t)$ reaches 94% success versus 12% for end-to-end RL and 38% for behavior cloning. The analysis of contact behavior shows the policy learns to switch contact faces to reduce estimator uncertainty, something the privileged policy applied to estimated states does less effectively (83% success).","pith_inferences":["The uncertainty estimates could be repurposed for failure prediction and fallback triggering, and for active data collection to improve estimator coverage; the authors list these as future work but the machinery already produces the quantities needed.","The synthetic occlusion model—freeze the last pose—is likely the tightest bottleneck; testing with partial or gradual occlusions on hardware would reveal whether the learned uncertainty calibration generalizes beyond that specific model.","The dependency on a fixed cuboid object and spherical pusher suggests the method's claims about uncertainty handling are demonstrated for one geometry; extending to diverse shapes would require re-collection of privileged data or a shape-conditioned estimator.","A direct comparison with a model-based filter (e.g., a particle filter fusing force and vision) on the same occlusion scenarios would clarify whether the learned uncertainty adds value beyond classical estimation."],"forward_implications":["Onboard-only perception becomes sufficient for planar non-prehensile manipulation under occlusion, removing the need for Vicon, under-table cameras, or multi-camera rigs.","Uncertainty-aware policies naturally learn information-gathering behaviors (e.g., switching contact faces) that reduce state uncertainty, improving task success beyond what nominal estimates provide.","Explicit state estimation with uncertainty beats both end-to-end policy learning and teacher-student behavior cloning for prolonged occlusion scenarios in this task.","Zero-shot sim-to-real transfer is possible even when the simulator omits pusher dynamics, provided observation noise is randomized aggressively during training."],"supporting_citations":[{"why":"Defines aleatoric and epistemic uncertainty and the negative log-likelihood loss used to train the estimator.","marker":"[30]"},{"why":"Provides the multivariate uncertainty formulation and the covariance combination used for total uncertainty.","marker":"[31]"},{"why":"Establishes MC dropout as approximate Bayesian inference, used for the epistemic uncertainty sampling.","marker":"[32]"},{"why":"PPO algorithm used to train both the privileged and uncertainty-aware control policies.","marker":"[35]"},{"why":"Dynamics randomization approach that motivates the zero-shot sim-to-real noise injection.","marker":"[7]"},{"why":"Prior planar pushing RL with multimodal categorical exploration supplying the action discretization and task setup.","marker":"[11]"},{"why":"GPU-accelerated physics simulation used to train policies and collect the interaction datasets.","marker":"[34]"},{"why":"Fiducial marker detection used to obtain object pose measurements on the hardware setup.","marker":"[18]"},{"why":"Teacher-student behavior cloning baseline that the paper compares against.","marker":"[27]"},{"why":"Dropout technique on which MC dropout sampling is built.","marker":"[33]"}],"fun_headline_variants":["Robots push unseen objects via tactile uncertainty learning","Occlusion-proof pushing with uncertainty-aware visuotactile policy","Tactile uncertainty beats vision for occluded object pushing","Learn to push blind: uncertainty-aware visuotactile control","Pushing what you can't see: uncertainty-aware estimator succeeds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic observation model used in training—occlusions that freeze the last observed pose, Gaussian noise levels in Table 1, and neglect of pusher dynamics compensated by large synthetic noise—represents the real robot's sensors and occlusion events closely enough that zero-shot sim-to-real transfer works.","fun_headline_variants_meta":{"raw":{"variants":["Robots push unseen objects via tactile uncertainty learning","Occlusion-proof pushing with uncertainty-aware visuotactile policy","Tactile uncertainty beats vision for occluded object pushing","Learn to push blind: uncertainty-aware visuotactile control","Pushing what you can't see: uncertainty-aware estimator succeeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1352,"prompt_tokens":1040,"completion_tokens":312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":656,"tokens_out":312,"duration_ms":3711,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:22:08.271517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Observe whether the estimated uncertainty rises during gradual partial occlusions on hardware; a failure to anticipate full occlusion would indicate the binary freeze-last-pose observation model is load-bearing.","supporting_citations":[{"cited_title":"Kendall and Y","cited_arxiv_id":null,"evidence_quote":"Defines aleatoric and epistemic uncertainty and the negative log-likelihood loss used to train the estimator."},{"cited_title":"Gal and Z","cited_arxiv_id":null,"evidence_quote":"Establishes MC dropout as approximate Bayesian inference, used for the epistemic uncertainty sampling."},{"cited_title":"Del Aguila Ferrandis, J","cited_arxiv_id":null,"evidence_quote":"Prior planar pushing RL with multimodal categorical exploration supplying the action discretization and task setup."},{"cited_title":"Wang and E","cited_arxiv_id":null,"evidence_quote":"Fiducial marker detection used to obtain object pose measurements on the hardware setup."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Teacher-student behavior cloning baseline that the paper compares against."},{"cited_title":"Srivastava, G","cited_arxiv_id":null,"evidence_quote":"Dropout technique on which MC dropout sampling is built."}],"review_version":1}