{"id":"a1876c09-4588-4a38-bd15-61f4ab264729","arxiv_id":"2504.12492","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A system estimates full-body pose and global translation in real time from one to three consumer-device IMUs, outperforming the prior IMUPoser approach on shared benchmarks.","lead":"MobilePoser tracks a person's full body and their movement through space using only the motion sensors already in their phone, smartwatch, or earbuds. It combines a neural network with a physics-based optimizer to run in real time on a phone, which could make full-body motion capture available to everyday fitness, gaming, and navigation apps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Translation evaluation protocol is internally inconsistent: Section 5.3.1 says fine-tuning on DIP-IMU, but Section 5.1 says DIP-IMU has no translation data, so the reported translation errors are not reproducible from the paper alone.","rationale":"The reader's weakest assumption was the fixed set of five body locations and zero-masking of absent devices. That is a real scope limitation, but the paper explicitly restricts itself to 24 combinations across those five locations and acknowledges drift in Section 8, so it does not invalidate the claimed configurations. The translation evaluation protocol is more load-bearing because global translation is the paper's primary new capability relative to IMUPoser, and Section 5.3.1 is internally inconsistent with Section 5.1. The reader noted this as a mechanical weakness but did not make it the central concern. I elevate it: if the translation numbers cannot be reproduced or assigned to a well-defined training protocol, the central claim that MobilePoser achieves accurate global translation from sparse consumer-device IMUs is not adequately supported. I considered whether the on-device real-time claim is also problematic, since the full pipeline is described as streaming to a MacBook and the iPhone measurement is only model-inference time, but that is secondary to the translation issue and the paper does present a CoreML conversion. The concern is substantial enough to keep the paper at CONDITIONAL with a request for a corrected protocol or code, but not severe enough to warrant rejection, because the ambiguity may be a reporting error and the pose-estimation contribution has independent support. Hence UNCHANGED.","tokens_in":14695,"tokens_out":6895,"duration_ms":74404,"concrete_test":"Obtain the released code or ask the authors for an explicit translation-training specification, then independently re-run the translation evaluation. Concretely: retrain v_e and v_f on AMASS only (no DIP-IMU fine-tuning), fuse them as described in Section 3.3.3, apply the physics optimizer, and evaluate Root Translation Error on TotalCapture and IMUPoser. If the reported 27.55 cm and 17.63 cm figures cannot be reproduced within a small tolerance, or if the authors confirm DIP-IMU was used to fine-tune translation networks despite lacking translation labels, the translation contribution loses its empirical support. If the numbers reproduce, the concern is resolved as a wording error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"MobilePoser's central claim has two parts: full-body pose and global translation. The pose side is supported by comparisons to IMUPoser on shared benchmarks. The translation side, which is the main new capability relative to IMUPoser, rests on Section 5.3.1. That section states: 'Like prior work, we use the Finetune DIP-IMU protocol, that is we train on AMASS and fine-tune on 8 participants of DIP-IMU to track the Root Translation Error.' But Section 5.1 explicitly says DIP-IMU 'does not contain global translation data.' Therefore the described protocol cannot be executed as written: the translation submodules (v_e, v_f) are trained with regression or classification losses requiring translation or foot-contact labels, and DIP-IMU provides neither. Two readings are possible: either only the pose model was fine-tuned on DIP-IMU while the translation module was trained on AMASS alone, or the translation module was somehow fine-tuned on DIP-IMU despite the missing labels. Both readings mean the reported translation errors of 27.55 cm (TotalCapture) and 17.63 cm (IMUPoser) are not reproducible from the manuscript and are not clearly comparable to prior work. This does not prove the translation claim is false, but it makes the central novelty unverifiable as written. The open-source link is unverified, so the protocol cannot currently be checked externally.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MobilePoser proposes a real-time system for full-body pose estimation and global 3D translation tracking using IMUs already present in consumer devices (smartphones, smartwatches, earbuds), with as few as one and up to three devices. The pipeline first estimates joint positions and rotations from masked IMU inputs using bi-LSTMs, then estimates root translation by fusing a foot-contact-based velocity estimate with a neural-network regressed root velocity, and finally applies a physics-based optimizer for temporal and physical plausibility. Evaluation is performed on DIP-IMU, TotalCapture, and IMUPoser datasets, with comparisons to IMUPoser and to 6-IMU systems such as TransPose and PIP. The manuscript also reports real-time on-device performance on an iPhone 15 Pro and describes proof-of-concept applications in indoor navigation, gaming, and fitness.","tokens_in":14981,"tokens_out":5304,"duration_ms":54448,"significance":"If the results hold, MobilePoser would be a meaningful advance for ubiquitous motion capture: it is, to my knowledge, the first system to demonstrate both full-body pose and global translation from 1-3 commodity IMUs, and it reports consistent pose-estimation improvements over IMUPoser on shared benchmarks, with a useful ablation showing that pose-aware translation regression reduces translation error. The real-time on-device implementation and the open-source release are also concrete strengths. However, the central novelty—global translation—is currently not verifiable from the manuscript because the evaluation protocol described in Section 5.3.1 is internally inconsistent with the dataset description in Section 5.1 and Table 2. The pose contribution is on firmer ground, but the paper's headline claim depends on the translation evaluation being corrected.","major_comments":[{"comment":"The translation evaluation protocol is not executable as written. Section 5.3.1 states that 'we train on AMASS and fine-tune on 8 participants of DIP-IMU to track the Root Translation Error,' but Section 5.1 explicitly states that DIP-IMU 'does not contain global translation data' and Table 2 marks DIP-IMU as lacking translation data. Since the translation submodules v_e and v_f are supervised with root velocities and foot-contact labels derived from translation data (Section 4.2), the described protocol cannot be carried out. Please specify exactly which submodules are fine-tuned on DIP-IMU and which labels are used. If the pose modules are fine-tuned on DIP-IMU while the translation modules remain trained on AMASS only, then the numbers 27.55 cm (TotalCapture) and 17.63 cm (IMUPoser) should be reported under a protocol that is clearly distinguished from the pose fine-tuning protocol, and the comparability to TransPose's fine-tuned translation results should be discussed. As written, the reported translation errors are not reproducible from the manuscript alone.","section":"Section 5.3.1, Section 5.1, Table 2"},{"comment":"Equation (6) as typeset is not a valid weighted average: the two fractions are identical, reading v = (q - q)/(q - q) * v_e + (q - q)/(q - q) * v_f. This would trivially reduce to v_e + v_f, not to a weighted blend. The surrounding text mentions an upper threshold and a lower threshold, and later gives q = 0.5 and q = 0.9, but the symbols in the equation are indistinguishable. Please introduce distinct notation, e.g., q_l and q_u, and typeset the formula correctly. Since this equation defines the core fusion mechanism for translation estimation, the intended form must be unambiguous.","section":"Equation (6)"},{"comment":"The claimed generality of the system is broader than what the method actually supports. The abstract and Section 1 claim estimation of pose and translation from 'any available subset of IMUs already present in these consumer devices' and 'various on-body configurations,' but Section 3.1 fixes the input to five predefined body locations (right/left pockets, right/left wrists, head) and restricts to 24 specific device-location combinations. Devices carried in unmodeled locations, such as a chest pocket, backpack, or belt clip, are outside the representation. Section 8 acknowledges drift when devices shift within clothing, but it does not acknowledge this representational restriction. Please temper the abstract and Section 1 claims, or explicitly state the supported set of locations and combinations in the contributions.","section":"Section 3.1, Section 1, Section 8"}],"minor_comments":[{"comment":"The text reads 'akin to prior wor' and should read 'akin to prior work.'","section":"Section 8"},{"comment":"The caption contains a duplicated article: 'shows the the cumulative distance error' should be 'shows the cumulative distance error.'","section":"Figure 8 caption"},{"comment":"The sentence 'This can be attributed to the legs capturing most of the locomotion data during translation' is slightly awkward; consider 'the legs capture most of the locomotion information.'","section":"Section 5.3.2"},{"comment":"Please clarify the deployment setup: the text first says the iPhone streams data to a MacBook Air 2022 via socket for inference, then later says a CoreML model is evaluated on an iPhone 15 Pro. It should be explicit whether the reported 60 fps real-time performance refers to the iPhone 15 Pro deployment only, and what role the laptop streaming prototype plays in the evaluation.","section":"Section 3.5"}],"recommendation":"major_revision","confidential_remarks":"The central novelty is global translation estimation, and the protocol inconsistency in Section 5.3.1 is the main barrier to acceptance. If the authors can clarify or correct the fine-tuning protocol, the contribution may be solid. The pose comparisons against IMUPoser are consistent and support the pose claim separately. I was unable to verify the GitHub repository during review; given the paper's reproducibility claims, the authors should ensure the link is live and contains the exact evaluation scripts. If the translation protocol cannot be corrected without changing the reported numbers, the claims should be scaled down accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Alex,\n\nQuick take: this paper earns a serious look. The new thing is real—full-body pose and global translation from one to three off-the-shelf IMUs (phone, watch, earbuds). IMUPoser does pose only; MobilePoser adds translation and shows consistent pose improvements on shared benchmarks. That is a meaningful step for HCI, health, and indoor navigation.\n\nWhat it does well: the pose side is credible. The comparisons with IMUPoser on DIP-IMU, TotalCapture, and the IMUPoser dataset show 12-24% MPJVE reductions across configurations. The ablations support the translation design (pose-aided velocity fusion helps, jerk loss and physics refinement reduce jitter). The system runs at 60 fps on an iPhone, which matters for real-world use. They also promise open-source code. All of that is solid system work.\n\nThe soft spots are mostly in the translation evaluation, which is the central novelty. Section 5.3.1 says they fine-tune on DIP-IMU for translation, but Section 5.1 explicitly says DIP-IMU has no translation data. That is a load-bearing inconsistency: the headline translation errors (27.55 cm on TotalCapture, 17.63 cm on IMUPoser) cannot be reproduced from the manuscript. Either they only fine-tuned the pose model and trained translation on AMASS alone, or the description is wrong—either way, the numbers as reported are not verifiable. That is not a fatal flaw in the approach, but it is exactly the kind of thing that has to be fixed before anyone should rely on the translation results. Equation (6) is also garbled, and there are no error bars anywhere, which makes the gains over IMUPoser look more certain than they are. The code link is unverified, so nothing can be checked externally yet.\n\nThe five fixed body locations (pockets, wrists, head) and T-pose calibration are a real limitation for in-the-wild use, but the authors acknowledge that in Section 8. I would not treat it as a fatal flaw for a first system paper.\n\nWho gets value from this: anyone working on inertial motion capture, ubiquitous sensing, or mobile health. It is a system paper with a clear benchmark story, not a theoretical breakthrough, but the combination is genuinely new and useful.\n\nMy recommendation: send it to peer review. The pose results are strong enough, and the translation claim is important enough to deserve referee time. But the reviewers should be told to dig into the translation protocol—it needs a clear rewrite and ideally error bars before acceptance.\n\nBest,\n[You]","headline":"A genuinely new capability—full-body pose plus global translation from 1-3 commodity IMUs—with solid pose gains over IMUPoser, but the translation evaluation protocol is internally inconsistent and the headline translation numbers are not reproducible from the paper alone.","tokens_in":15561,"tokens_out":1237,"would_cite":true,"duration_ms":14281,"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":"A phone, watch, or earbuds alone can drive real-time full-body pose and walking-path tracking, with no cameras or mocap suit.","keywords":["full-body pose estimation","inertial measurement units","global translation","consumer devices","sparse sensing","deep neural network","physics-based optimization","real-time"],"falsifier":"Run MobilePoser's trained model with a phone carried in a left chest pocket, a backpack strap, or an ankle holster — locations outside the five predefined sites — while a reference optical mocap system records ground truth. If per-joint vertex error or root translation error degrades substantially compared to the same user wearing the phone in a supported location, or if the model fails to track at all, the claim that 'any available subset of IMUs already present in these consumer devices' suffices is falsified; the paper's own stated drift sensitivity when devices deviate from calibrated positions makes this the direct check.","tokens_in":14447,"feed_emoji":"🏃","tokens_out":10271,"duration_ms":91070,"temperature":0.7,"pith_summary":"This paper argues that the inertial sensors already inside consumer devices — a phone, a smartwatch, or earbuds — carry enough information to reconstruct a person's full-body pose and their movement through space in real time, using as few as one device and no more than three. The proposed system, MobilePoser, estimates joint positions and then joint rotations from sparse, partially absent IMU streams with a multi-stage neural network, then passes the result through a physics-based optimizer that removes jitter, foot sliding, and floor penetration. Global translation is obtained by fusing a foot-ground-contact velocity estimator with a direct neural regressor of root velocity, a combination that handles both walking and aerial phases like running or jumping. Across the DIP-IMU, TotalCapture, and IMUPoser benchmarks, the paper reports pose errors 8.7–24.1% lower than the closest consumer-device baseline (IMUPoser) and a mean root translation error of 17.63 cm on consumer-device data, at 60 frames per second on an iPhone 15 Pro. If the claims hold, motion capture becomes an infrastructure-free capability of devices people already own, relevant to fitness, gaming, and indoor navigation.","feed_headline":"A phone in your pocket can now track both pose and walking path","feed_subtitle":"No cameras, no motion-capture suit: everyday devices estimate the body and its movement in real time.","key_machinery":"The central mechanism is a zero-masked input representation over five fixed body locations — right and left pockets, right and left wrists, head — combined with a two-stage pose network and a fused translation estimator. The input vector $\\boldsymbol{x} \\in \\mathbb{R}^{60}$ (acceleration and orientation for each of five locations) has zeros substituted for absent devices, so a single unified model handles any 1–3 device subset without retraining per combination. Pose estimation proceeds from joint positions (chosen because positions correlate linearly with acceleration) to joint rotations (regressed from positions plus IMU data), with a jerk loss enforcing temporal smoothness. Translation estimation is the load-bearing fusion: the foot-ground-contact submodule tracks the supporting foot's displacement, which is reliable during stance but fails in flight phases, so a direct LSTM regressor of root velocity covers the gaps, and a confidence-weighted sum of the two produces the final global velocity. A physics-based optimizer from prior work [48] is the final stage, refining both pose and translation to satisfy the equation of motion and reduce jitter and foot-floor penetration.","core_discovery":"On the paper's own terms, the discovery is that low-fidelity, heterogeneous, and extremely sparse IMUs — one to three streams coming from phones, watches, or earbuds — are enough for real-time full-body pose and global translation estimation, a problem previously requiring six or more specialized sensors. The system's design is multi-stage: a bi-LSTM predicts 3D joint positions (an intermediate representation with linear correlation to acceleration), a second bi-LSTM regresses 18 joint rotations from those positions plus raw IMU data, and a translation module fuses foot-ground-contact-derived root velocity with a direct LSTM root-velocity regressor using a confidence-weighted sum. A physics-aware optimizer from prior work [48] then enforces physical plausibility, reducing jitter and ground penetration. The reported results — lower vertex error than IMUPoser across all device counts on the IMUPoser dataset, and translation errors comparable to some six-IMU systems at a fraction of the instrumentation — support the paper's positioning of MobilePoser as a practical, on-device digitization technique. The system also demonstrates that higher-order digitization (full-body pose) can be used to improve lower-order digitization (steps and translation), a key architectural insight of the ablation study.","pith_inferences":["Editorial inference: The first real-world failure mode is probably the five-location assumption: if a phone rides in a chest pocket, backpack, or belt clip, or shifts inside loose clothing after calibration, the zero-masked representation silently encodes the wrong body site, and the paper's acknowledged drift sensitivity implies the model will hallucinate poses rather than signal the error.","Editorial inference: The physics optimizer's residual could be recycled as an online self-diagnostic: an increasing penalty over time would indicate that the device configuration no longer matches the calibration, enabling automatic re-calibration without user intervention.","Editorial inference: Since the paper's evaluations are all on lab-collected datasets, the natural next test is an in-the-wild benchmark with a portable optical ground-truth system; the current evidence does not establish how the method behaves under the full range of real-world device malpositioning.","Editorial inference: The 'higher-order digitization helps lower-order digitization' insight could extend to other downstream tasks like fall detection or gait analysis, where a pose prior might regularize noisy raw IMU features in the same way it improves translation."],"forward_implications":["Fitness and wellness apps can offer real-time exercise-form feedback, rep counting, and calorie estimates using only a phone in a pocket or a watch on a wrist, with no camera or lab setup.","Indoor navigation and localization can be driven by the same IMU streams, giving walking path and direction without GPS or fixed beacons.","Immersive mobile gaming and VR/AR avatars can be controlled by full-body motion from devices the player already wears, removing external sensors and specialized controllers.","Because the pose network is trained on large synthetic motion-capture data and fine-tunes to different IMU noise profiles, the method can adapt to new device placements and activities without collecting new paired sensor data.","The 60 fps on-device performance means the whole pipeline can be embedded in consumer apps, enabling passive, continuous user digitization in everyday settings."],"supporting_citations":[{"why":"Defines the five body locations and 24 device-location combinations MobilePoser adopts, the active-device selection strategy, and the baseline plus dataset for the central pose comparisons.","marker":"[28]"},{"why":"Supplies the foot-ground-contact-based root velocity estimation, the confidence-weighted fusion heuristic, and the jerk loss used for temporal smoothness.","marker":"[49]"},{"why":"Provides the physics-aware motion optimizer that refines joint rotations and translation to reduce jitter, foot skating, and floor penetration.","marker":"[48]"},{"why":"Supplies the method for synthesizing IMU measurements from virtual sensors on the mesh, the bi-LSTM architecture, and the DIP-IMU benchmark for evaluation.","marker":"[14]"},{"why":"Provides the AMASS motion-capture corpus used to generate the synthetic training data for pose and translation networks.","marker":"[26]"},{"why":"Defines the SMPL body model (24 joints, 6890-vertex mesh) that serves as the pose representation and the target for generated meshes.","marker":"[25]"},{"why":"Supplies TotalCapture, a real-IMU dataset with ground-truth translation, used for the translation evaluation protocol.","marker":"[40]"}],"fun_headline_variants":["Phone IMUs alone track full-body pose and walking path","No specialized sensors: consumer IMUs do real-time mocap","Your phone, watch, or earbuds can now capture full-body motion","Sparse IMUs from everyday devices enable real-time pose tracking","Full-body motion capture from just a few consumer-device IMUs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system assumes that a user's devices are always worn or carried at one of exactly five fixed body locations — right and left pockets, right and left wrists, and head — and that a T-pose calibration plus zero-masking of absent devices fully captures how consumers carry them; if a device rides somewhere else, such as a chest pocket, backpack, or belt clip, or shifts within clothing after calibration, the input representation and calibration are no longer valid.","fun_headline_variants_meta":{"raw":{"variants":["Phone IMUs alone track full-body pose and walking path","No specialized sensors: consumer IMUs do real-time mocap","Your phone, watch, or earbuds can now capture full-body motion","Sparse IMUs from everyday devices enable real-time pose tracking","Full-body motion capture from just a few consumer-device IMUs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":1165,"prompt_tokens":972,"completion_tokens":193,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":104}},"tokens_in":588,"tokens_out":193,"duration_ms":2967,"temperature":1.0,"reasoning_tokens":104,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:30:34.813804+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MobilePoser's trained model with a phone carried in a left chest pocket, a backpack strap, or an ankle holster — locations outside the five predefined sites — while a reference optical mocap system records ground truth. If per-joint vertex error or root translation error degrades substantially compared to the same user wearing the phone in a supported location, or if the model fails to track at all, the claim that 'any available subset of IMUs already present in these consumer devices' suffices is falsified; the paper's own stated drift sensitivity when devices deviate from calibrated positions makes this the direct check.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the foot-ground-contact-based root velocity estimation, the confidence-weighted fusion heuristic, and the jerk loss used for temporal smoothness."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the physics-aware motion optimizer that refines joint rotations and translation to reduce jitter, foot skating, and floor penetration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies TotalCapture, a real-IMU dataset with ground-truth translation, used for the translation evaluation protocol."}],"review_version":1}