{"id":"cffc1a0e-3098-4e40-bc09-2c8887c3d302","arxiv_id":"2504.18792","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A real-time action correction system, STDArm, transfers visuomotor policies trained on static data to moving platforms, recovering 40 to 93 percent of static success rates in three manipulation tasks without retraining the policy.","lead":"STDArm is a software layer that lets robot manipulation policies trained on static, table-mounted arms run on moving platforms such as drones and wheeled tables, by correcting each action in real time for platform motion. A generalist would read it as a practical alternative to recollecting training data for every new mobile robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline dynamic results rely on uncontrolled manual shaking; without matched disturbance conditions, reported gains may reflect trial difficulty rather than STDArm.","rationale":"The reader's weakest assumption—that the foundational policy remains functional under motion—is a genuine limitation and is explicitly acknowledged in the paper. However, it does not attack the internal correctness of the action-correction mechanism; it only restricts the claim's scope. The more load-bearing problem is that the headline dynamic results were obtained without controlled, reproducible disturbance conditions. The manual-shaking protocol, even if double-blind, does not guarantee that baseline and STDArm trials face comparable platform motion, and the absence of confidence intervals over 15-trial runs leaves the reported improvements vulnerable to trial-to-trial variability. The LS-3DoF setup, which would provide reproducible motion, is used only for ablations, so the central baseline-versus-STDArm comparison lacks the experimental control needed to attribute the gains to the method. The UAV result adds a second confound: the human operator manually maintains relative positioning, so the demonstration is not fully autonomous. My recommendation remains conditional: the method is plausible and the ablations are suggestive, but the central quantitative claim should be tested under matched, measurable disturbance conditions before it is accepted as established. This does not move the reader's verdict because the reader already called for more trials and clearer experimental evidence, but I identify the experimental-control issue as the specific decisive check.","tokens_in":12726,"tokens_out":7955,"duration_ms":83657,"concrete_test":"Re-run Table I's three tasks on a motorized, reproducible platform (e.g., LS-3DoF or a linear stage) with prerecorded disturbance trajectories of matched amplitude and frequency, executing both the base policy and STDArm on the exact same trajectory set for at least 50 trials per condition. Report success rates with 95% confidence intervals and per-axis platform motion (e.g., T265/IMU pose RMS). If STDArm's advantage persists under matched disturbances, the confounding concern is resolved; if not, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that STDArm restores static policy performance under platform motion—is tested almost entirely under manually generated, uncontrolled shaking. In Sec. V-A, MT-3DoF and MF-5DoF use an operator shaking the platform; the protocol is double-blind, but no measurement or recording of the disturbance is reported, so baseline and STDArm trials cannot be shown to face comparable motion conditions. With 15 trials per cell and no confidence intervals, a run-to-run difference in shaking amplitude or frequency could account for part of the gap (e.g., 0% vs 53.3% on Pick and Place). The one reproducible platform, LS-3DoF, is used only for ablation (Sec. V-C), not for the headline baseline-vs-STDArm comparison. The UAV experiment is further confounded: an operator manually maintains the drone's relative position to the target, so the 40% success rate does not demonstrate autonomous compensation for platform drift. The acknowledged perception limitation (Limitations section) is real but narrower: it caps scope, whereas the uncontrolled-disturbance issue affects the quantitative evidence for the core claim. A direct controlled test is needed before the reported gains can be attributed to STDArm rather than to trial difficulty.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents STDArm, a modular system for transferring visuomotor policies trained under static conditions to mobile robotic platforms. The system consists of three components: an action manager that performs temporal ensembling and interpolation to raise control frequency, a prediction-network-based stabilizer that compensates for platform motion using high-frequency visual SLAM poses, and an online latency estimation module that calibrates system delay during a pre-mission warm-up. The authors evaluate STDArm on four platform configurations (mobile table, mobile frame, leadscrew, drone), two policies (Diffusion Policy and ACT), and three manipulation tasks, reporting that STDArm substantially improves success rates under dynamic conditions and preserves static performance. The core claim is that STDArm enables real-time motion compensation with centimeter-level precision without modifying the foundational policy.","tokens_in":12950,"tokens_out":3794,"duration_ms":40718,"significance":"The problem addressed is timely and practical: static-trained visuomotor policies are difficult to deploy on legged, wheeled, and aerial platforms because of platform motion, inference latency, and limited onboard compute. STDArm's design is appealing in its modularity: it does not require retraining the policy, works with multiple policy architectures, and runs entirely on edge hardware. If the reported gains are reproducible, the system would be a useful and low-cost tool for mobile manipulation. The experiments span multiple platforms, arms, and tasks, and the authors include ablations on a programmable leadscrew platform to control for some sources of variability. The paper also provides a project page and builds on public codebases. However, the central quantitative claim is currently supported only by a small number of trials under manually generated and unmeasured disturbances, which makes it difficult to attribute the observed improvements to STDArm specifically. The latency-estimation procedure also has a circularity issue in one of the demonstration tasks.","major_comments":[{"comment":"The headline dynamic results for MT-3DoF and MF-5DoF rely on manually shaking the platform, but the disturbance is neither recorded nor matched between the baseline and STDArm conditions. With only 15 trials per cell, a difference in shaking amplitude or frequency between runs could explain part of the reported gap (for example, Pick and Place on MT-3DoF from 0% to 53.3%). The double-blind protocol is commendable, but it does not guarantee that the baseline and STDArm trials face equivalent disturbances. The reproducible LS-3DoF platform is used only for ablations, not for the headline baseline-versus-STDArm comparison. A controlled or at least recorded-disturbance experiment is needed before the gains in Table I can be attributed to STDArm rather than to trial difficulty.","section":"Sec. V-A (Experimental Setup) and Table I"},{"comment":"Every reported success rate is based on 15 trials, with no error bars, confidence intervals, or significance tests. Several differences are within a few trials: for example, 66.7% vs. 53.3% in Stack Cylinder on MF-5DoF, and 40% vs. 13.3% on UAV-3DoF. Without statistical treatment or a larger number of trials, the claimed 'recovery' of static-level performance is not firmly established. At minimum, exact binomial confidence intervals should be reported, and the authors should state how many trials were performed per cell and whether trials were randomized in blocks.","section":"Table I and Fig. 7"},{"comment":"The system latency is calibrated online by minimizing the red marker's pixel-plane movement in an end-hold task (Sec. III-D), and the same end-hold task and the same marker-movement metric are then used in Sec. V-D to demonstrate the stabilizer's benefit (Fig. 8). This is partly circular: the configuration is tuned to optimize exactly the metric used for evaluation. The demonstration would be more convincing if it used an independent metric, such as measured end-effector position error under held-out motion conditions, or if the end-hold task success was compared against a separate task objective.","section":"Sec. III-D and Sec. V-D (Fig. 8)"},{"comment":"For the UAV experiment, the text states that 'an operator manually maintains the relative positioning between the drone and the delivery target.' This means the 40% success rate with STDArm does not demonstrate autonomous compensation for platform drift; it shows only that STDArm can help under human-assisted station-keeping. The related claim in Sec. V-B that the UAV result highlights 'adaptability to ... inherent aerodynamic uncertainties' is therefore overstated. To support the aerial-platform claim, the experiment would need to quantify the platform's position error and separate autonomous stabilization from operator assistance.","section":"Sec. V-A (UAV-3DoF) and Table I"},{"comment":"The paper repeatedly claims 'centimeter-level operational precision' and 'centimeter-level precision,' but no direct precision measurement is reported. The tasks have tolerance specifications (e.g., 4 cm for stacking), but success rates are not equivalent to a precision measurement. The end-hold heatmaps in Fig. 8 are qualitative. To support this claim, the authors should report actual end-effector error statistics (mean, std, or a percent-within-threshold) under dynamic conditions.","section":"Abstract and Sec. VI"}],"minor_comments":[{"comment":"The action interpolation formula appears to have its weights reversed: at τ = t it gives A_{t+1} instead of A_t. As written, A_τ = (τ−t)A_t + (t+1−τ)A_{t+1}; the intended formula is likely A_τ = (t+1−τ)A_t + (τ−t)A_{t+1}. Please correct the equation or clarify the definition of τ.","section":"Eq. (6)"},{"comment":"The axis labels in Fig. 7 are garbled/unreadable in the provided version; the figure should be regenerated with clear, legible labels.","section":"Fig. 7"},{"comment":"The definition of Δt_τ is not explicitly distinguished from Δt elsewhere in the text; please clarify that Δt_τ is the interval from action generation to execution for action Aτ, while Δt is the total latency estimated online.","section":"Sec. III-C, Eq. (9)"},{"comment":"The training of the pose prediction network is described only as 'approximately five minutes of motion data.' Please provide details on the supervision signal and how future poses are obtained for training, as this is central to the stabilizer's operation.","section":"Sec. V-A"},{"comment":"The abstract states the SLAM operates at '10× the base policy's observation rate.' With SLAM at 200 Hz and the DP policy at 5 Hz, the factor is 40×; for ACT at 30 Hz it is about 6.7×. Please make this statement consistent with the actual frequencies.","section":"Intro, Sec. I"},{"comment":"There are several typographical artifacts, such as 'UA V-3DoF' with an extra space and inconsistent formatting in the platform names; these should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real deployment problem and the proposed system is plausible, but the current evidence base is too weak for the strength of the claims. The uncontrolled manual shaking in the primary evaluation, the lack of statistical reporting, and the circularity in the end-hold demonstration all need to be addressed. I would like to see the authors either use the programmable leadscrew as the primary controlled condition, or record and characterize the manual disturbances so that baseline and STDArm trials are matched. The 'centimeter-level precision' claim should be backed by direct error metrics. I do not think rejection is warranted, because the central idea is sound and the limitations paragraph honestly acknowledges the policy-generalization scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful systems paper, not a breakthrough. The action manager + stabilizer + online latency estimator is a genuinely practical wrapper: you keep your static-trained DP or ACT policy, add high-frequency pose readout and action correction, and it works on moving bases without retraining. That is a real contribution over UMI on Legs (which needs mobile data) or MobileALOHA (static-only). The tests span two arms, two policies, three tasks, four platforms including a drone, and the whole thing runs on a Jetson Xavier. The ablations show each component earns its keep, especially latency estimation. The limitations section is admirably honest: if the base policy's perception fails under motion, this wrapper can't save it.\n\nWhere it's soft: the headline evidence is 15 trials per cell with no error bars or significance tests. With success rates 0 vs 53.3%, a few trials move the number a lot. More serious, the dynamic conditions on MT-3DoF and MF-5DoF come from an operator shaking the platform; the protocol is double-blind, which should prevent bias, but it doesn't equalize disturbance across runs. The one reproducible platform (LS-3DoF, leadscrew) is only used for ablation, not for the headline baseline-vs-STDArm comparison. That was a missed opportunity: a controlled disturbance on the main task would have made the case much stronger. The UAV experiment is fine as scoped—the operator keeps relative position, so it's about compensation for hovering instability, not autonomous flight—but again 15 trials and a 40 vs 13.3% gap.\n\nTwo smaller issues: the end-hold demonstration (Fig 8) uses the same marker-minimization criterion that was used for latency calibration, so that particular plot is partly circular; the main manipulation experiments don't suffer from this. And the 'centimeter-level precision' claim is inferred from task tolerances, not measured directly; should be softened.\n\nVerdict: the central idea holds up, the evidence is thinner than advertised. I'd send it to peer review with a request for more trials, a controlled disturbance condition on a headline task, and direct precision measurements. Release of code/data would help too. The paper deserves a serious referee, but needs revision before acceptance.","headline":"A practical wrapper for deploying static-trained policies on moving robots, with consistent gains across platforms, but the headline experiments rely on 15 trials with manually shaken bases and no controlled disturbance, so the quantitative claims are indicative rather than proven.","tokens_in":13478,"tokens_out":3175,"would_cite":true,"duration_ms":33847,"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":"STDArm shows that static-trained visuomotor policies can transfer to moving robot platforms through real-time action correction, recovering near-static success rates without retraining.","keywords":["STDArm","visuomotor policies","imitation learning","mobile manipulation","action correction","pose prediction","latency compensation","drone manipulation"],"falsifier":"Record success with STDArm active while moving the base to viewpoints or lighting conditions not present in the static training data, for example rotating the camera to look at the object from above or swapping the background. If success drops to the uncorrected baseline level, the claim that STDArm preserves manipulation capability under platform motion is disproved for those conditions.","tokens_in":12539,"feed_emoji":"🤖","tokens_out":11367,"duration_ms":94146,"temperature":0.7,"pith_summary":"STDArm is a system for moving imitation-learned visuomotor policies from fixed tabletops to moving platforms, including a drone. The paper's central claim is that a real-time action-correction layer, rather than retraining or recollecting demonstrations, is enough to recover near-static manipulation performance while the base moves. This matters because high-quality demonstrations are far easier to collect on a fixed arm, and existing mobile-manipulation pipelines either require specialized data collection or accept large drops in success. In experiments with diffusion policy and ACT on three tasks, STDArm raises dynamic success rates from near zero to values close to the static baselines. The system runs entirely on an edge computing board.","feed_headline":"Static-trained robot policies keep working while the robot moves","feed_subtitle":"STDArm lifts mobile success rates from near zero to static-level performance without retraining.","key_machinery":"The mechanism that carries the argument is the stabilizer's action-correction transform. At each high-frequency control step $\\tau$, with total system latency $\\Delta t$, extrinsic calibration $E$ between the SLAM frame and the arm frame, and platform pose stream at frequency $f$, the stabilizer forms $\\delta(\\tau,\\Delta t,E) = E^{-1}\\left(\\Delta p_{\\lceil -\\Delta t \\tau/f \\rceil}^{-1}\\,\\Delta p^*_{\\lfloor \\Delta t/f \\rfloor}\\right)^{-1}E$, where $\\Delta p$ is an observed past relative pose and $\\Delta p^*$ a predicted future relative pose from the lightweight network. The final command is $A'_\\tau = \\delta(\\tau,\\Delta t,E)\\, A_\\tau$, so the arm's motion relative to the moving base is what the policy intended relative to the world. This correction works only because the action manager first provides a dense 50 Hz action stream by temporal ensembling overlapping action chunks with exponentially decaying weights $w_{t+i} = \\exp(-\\alpha i)$ and linearly interpolating between buffer entries; the prediction network itself is a parallel long short-term memory (LSTM) and gated recurrent unit (GRU) whose concatenated outputs feed a fully connected layer.","core_discovery":"On the paper's own terms, STDArm's discovery is that a static-trained visuomotor policy can be transplanted to a moving platform by treating deployment as a latency and frequency alignment problem rather than a policy-learning problem. The policy network is untouched; instead, an action manager raises and stabilizes the action stream, a stabilizer predicts the platform's short-term motion with a lightweight LSTM-GRU network and applies the correction $\\delta(\\tau,\\Delta t,E)$ to every action, and a warm-up routine calibrates the total system latency online. The quantitative claim is that this restores near-static success: DP+STDArm on the mobile-table arm goes from 0% to 53.3% on Pick and Place and from 20% to 93.3% on Get the Delivery, and the drone version of Get the Delivery rises from 13.3% to 40%, while static performance is preserved.","pith_inferences":["An extension the paper suggests but does not demonstrate is that any action-chunked policy outputting end-effector poses could be wrapped by the same layer, since the policy network's weights are never modified.","Because the correction depends on a visual-inertial SLAM pose stream, the system will inherit SLAM drift or tracking loss in low-texture or outdoor environments; the paper does not evaluate these failure modes.","The pre-mission latency search could be replaced by continuous online calibration, which would track changing compute loads on shared edge devices; this is a testable improvement rather than a paper claim.","The LS-3DoF ablation shows stronger gains on regular, predictable motion, suggesting the predictor exploits predictability; isolating motion predictability as an independent variable would separate this effect from pure latency compensation."],"forward_implications":["Static demonstration datasets, which are cheap to collect on fixed arms, become usable for drones, quadrupeds, and other moving bases without recollecting data on the moving platform.","Action-chunked visuomotor policies can be deployed at higher effective control frequency on low-cost edge hardware, improving responsiveness without changing the policy network.","Ablations on the leadscrew setup show the components are complementary: the full STDArm reaches 80% on Get the Delivery, while removing online latency estimation or the action manager drops success to 53.3% or below.","The drone result, though modest at 40% versus 13.3%, suggests aerial manipulation is within reach of static-trained policies once motion compensation and latency are handled."],"supporting_citations":[{"why":"Diffusion Policy is the main wrapped policy: its 20-step iterative denoising at low frequency creates the action-chunking latency that STDArm must compensate.","marker":"[4]"},{"why":"ACT is the second wrapped policy and the source of the temporal-ensemble smoothing the action manager uses.","marker":"[3]"},{"why":"UMI on Legs is the previous approach for deploying diffusion policies on legged platforms; STDArm extends this to more dynamic bases.","marker":"[7]"},{"why":"Mobile ALOHA is the wheeled mobile manipulation baseline whose static-operation assumption STDArm's design targets.","marker":"[17]"},{"why":"LSTM is one of the two parallel recurrent networks in the stabilizer's pose predictor.","marker":"[32]"},{"why":"GRU is the other parallel recurrent network in the stabilizer's pose predictor.","marker":"[33]"},{"why":"ResNet-18 is the vision backbone of both policy networks in the experiments, carrying the perception on which the whole pipeline depends.","marker":"[34]"}],"fun_headline_variants":["Static-trained policies go mobile without retraining","STDArm: centimeter precision on moving robots","No retraining: static policy works on moving platform","Motion-proof policies: static training, dynamic success","Moving robots keep static-trained skills intact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The visuomotor policy must still recognize the target from images; STDArm only corrects actions, so if the moving camera's changing viewpoint or background defeats the policy's perception, the system cannot restore the task.","fun_headline_variants_meta":{"raw":{"variants":["Static-trained policies go mobile without retraining","STDArm: centimeter precision on moving robots","No retraining: static policy works on moving platform","Motion-proof policies: static training, dynamic success","Moving robots keep static-trained skills intact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1466,"prompt_tokens":928,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":544,"tokens_out":538,"duration_ms":5304,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:08:56.126972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record success with STDArm active while moving the base to viewpoints or lighting conditions not present in the static training data, for example rotating the camera to look at the object from above or swapping the background. If success drops to the uncorrected baseline level, the claim that STDArm preserves manipulation capability under platform motion is disproved for those conditions.","supporting_citations":[{"cited_title":"Learning fine-grained bimanual manipulation with low-cost hardware","cited_arxiv_id":null,"evidence_quote":"ACT is the second wrapped policy and the source of the temporal-ensemble smoothing the action manager uses."},{"cited_title":"Long short-term memory","cited_arxiv_id":null,"evidence_quote":"LSTM is one of the two parallel recurrent networks in the stabilizer's pose predictor."}],"review_version":1}