{"id":"e150a6ba-9c0d-43db-bfe4-e68259070891","arxiv_id":"2505.09780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Replacing raw IMU inputs with level-crossing events on SE(3) pre-integrations improves generalization and accuracy of neural displacement priors for inertial odometry.","lead":"A robotics team trained inertial navigation networks on 'Lie events' that fire only when the IMU pre-integration on the SE(3) group changes by a fixed amount, rather than on fixed-rate IMU samples. This representation improves the robustness of standard neural inertial odometry to sampling rate and speed changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's invariance assumes the pre-integration signal factorizes as x(t)=x*(φ(t)); with estimated initial velocity v0 the pre-integrated translation contains a v0-error drift term, so event timestamps and polarities are no longer path-only.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Theorem 1 assumes a perfect factorization x(t)=x*(φ(t)), which is violated by pre-integration drift, most directly from the estimated initial velocity v0. I agree this is the central soft spot. The paper's own limitations section admits pre-integration drift, and the toy validation does not include v0 perturbations, so the theory-to-practice gap is real and untested. Other potential concerns (e.g., no error bars, partial canonicalization of raw ω/a inputs, or the fact that the theoretical result covers only speed reparametrizations and not arbitrary trajectory shape changes) are secondary: they affect the strength or interpretation of the empirical claims, but the v0-dependence strikes at the core of the claimed invariance. A focused experiment measuring event-timestamp and polarity fidelity under v0 perturbation would settle whether the concern lands. If the degradation is small, the conditional acceptance stands; if it is large, the paper's theoretical claim would need to be revised or the method's design changed. Since the reader already judged the paper CONDITIONAL and this concern supports that judgment, I recommend no change to the verdict.","tokens_in":23942,"tokens_out":8837,"duration_ms":91570,"concrete_test":"Use the Sec. IV-A toy setup, but generate events with the true initial velocity v0* and with perturbed velocities v0*+δ for δ∈{−0.5,0,0.5} m/s, across α∈{0.5,2.0} and θ∈{0.005,0.01,0.02}. Compute (i) the chamfer distance between φ_α(E_δ) and E* exactly as in Table I, and (ii) the mean angular error between the Lie polarities p_δ(τ_j) and the canonical p*(σ*_j) after matching events by arc-length index. If both metrics remain within 2× of the δ=0 baseline, the v0 sensitivity is moderate and the core invariance is approximately preserved; if the chamfer distance or polarity error degrades by more than 10× or by >10°, the theoretical canonicalization does not survive realistic v0 uncertainty, and the paper should either restrict its invariance claim to the idealized setting or modify the reference signal to remove the v0 dependence (e.g., using relative displacement deltas).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim rests on Theorem 1, which requires the reference signal to satisfy x(t)=x*(φ(t)) exactly. In the actual pipeline, the reference signal is the IMU pre-integration (Sec. III-C), whose translation depends on the initial velocity v0 through Eqs. (19)-(20). At test time v0 is taken from the EKF clone state; at training it is ground truth plus uniform noise in [-0.5,0.5] m/s. With an initial-velocity error Δv, the pre-integrated position evolves as t̃(t)=t*(φ(t))+Δv(t−t1), which cannot be written as t̃*(φ(t)) for a fixed path t̃* unless φ is linear. Consequently, the induction step of Theorem 1 that relies on x_ref,j=x*_ref,j fails: both event timestamps and Lie polarities acquire a dependence on the unobservable v0 error, and the claimed path-only canonicalization is broken. Equation (22) asserts the factorization only 'approximately,' without an error bound, and the Limitations section (Sec. V) concedes that pre-integrations 'suffer from slight drift and sampling noise.' The toy experiment in Table I does not perturb v0, so it cannot validate the invariance under the actual operating conditions. The end-to-end sensitivity analysis (Fig. 6) shows only graceful degradation of ATE*, not whether the event-generation invariance itself survives. Thus the paper's theoretical support for the central claim is not established for the real pre-integration pipeline, and the empirical gains may arise from the network learning to cope with v0-dependent events rather than from true canonicalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing raw fixed-rate IMU samples with 'Lie events' as inputs to neural displacement priors (NDPs) for inertial odometry. Events are triggered whenever the norm of the change in the SE(3) IMU pre-integration, measured through the Lie algebra, exceeds a contrast threshold θ. Each event carries the IMU measurements at that time and a 'Lie polarity', a normalized element of the Lie algebra encoding the direction of change. The authors state a theoretical invariance result (Theorem 1): if the reference signal factorizes as x(t)=x*(φ(t)) with φ'(t)>0, then event timestamps transform as φ^{-1} and reference poses are independent of φ. They apply the method to TLIO and RoNIN across multiple datasets, reporting reduced ATE/MSE and improved robustness to IMU sampling rate without rate augmentation, and provide open-source code.","tokens_in":24268,"tokens_out":11429,"duration_ms":114713,"significance":"If the claimed invariance holds in practice, the method is a simple, broadly applicable preprocessing that could improve the generalization of neural inertial odometry to different speeds and sampling rates. The generalization of event-based level-crossing and binary polarities to SE(3) is a novel formalism that may be useful beyond odometry. The experimental breadth is a strength: results are reported for TLIO, Aria, RoNIN, RIDI, and OxIOD, with two different NDP architectures, and rate robustness is demonstrated without rate augmentation. However, the theoretical guarantee is proven only under an exact trajectory factorization that the actual pre-integration pipeline does not satisfy, and the headline error reductions are reported without uncertainty estimates; the paper's central claim therefore currently rests on a gap between the theorem's assumptions and the deployed system.","major_comments":[{"comment":"The proof of Theorem 1/Theorem 2 in the appendix is technically incorrect for the 1D case: the event definition in Eq. (15) uses the condition θ ≤ ||log x(t) − log x_ref,j−1||, but the proof in the appendix uses θ ≤ ||x(t) − x_ref,j−1||. These conditions are not equivalent, so the proof as written does not establish the theorem for the definition actually used. The argument can likely be repaired by carrying the logarithm through the reparametrization, but the present proof needs correction.","section":"Section III-B and Appendix VI-A"},{"comment":"The factorization x(t)=x*(φ(t)) that Theorem 1 requires is not satisfied by the IMU pre-integration used as the reference signal in the actual pipeline. The pre-integration recursion (19)–(20) depends on the initial velocity v0, which at test time is an EKF estimate and at training is a perturbed ground-truth value. An initial-velocity error Δv introduces a term Δv(t−t1) into the pre-integrated position, which cannot be written as a fixed path under a time reparametrization unless φ is affine. Equation (22) asserts the factorization only 'approximately', with no error bound, and Section V concedes that pre-integrations 'suffer from slight drift'. Consequently, event timestamps and Lie polarities are not path-only in the deployed system, so the theoretical canonicalization claim is not established for the real pipeline.","section":"Section III-C, Eqs. (18)–(22)"},{"comment":"The toy experiment in Table I does not perturb v0, so it cannot validate the invariance under the actual operating conditions described in Sections IV-B and IV-C. The end-to-end sensitivity study in Fig. 6 shows graceful degradation of ATE* under v0 noise, but it measures only the final trajectory error, not the event timestamps or polarities, so it does not confirm whether the event-generation invariance itself survives. The theoretical claim and the empirical validation therefore address different quantities.","section":"Section IV-A, Table I and Fig. 6"},{"comment":"The headline reductions (e.g., 'up to 21%' in the abstract) are reported as single median values over the test set, without error bars, confidence intervals, or multiple training runs. Since the central empirical claim is that Lie events reduce downstream odometry error, single-point medians are insufficient to assess statistical significance, and the percentages should be accompanied by a measure of spread (e.g., per-sequence interquartile ranges, or mean±std over several independent training runs).","section":"Tables II and III"},{"comment":"The theoretical invariance applies only to event timestamps and Lie polarities; the network inputs also include a(τ_j) and ω(τ_j), which per Eqs. (12)–(13) still depend on φ̇ and φ̈. The abstract's phrase that events 'are robust to input rate changes and have favorable invariances' therefore overstates the scope of the theoretical result. The empirical rate-robustness experiments are the actual support for the broader claim, and the paper should present the theoretical result as applying to timestamps and polarities only, with the end-to-end rate robustness established empirically.","section":"Section III-A, Eqs. (12)–(14)"}],"minor_comments":[{"comment":"In the base step of the proof, 'τ_j = φ^{-1}(τ_j) = φ^{-1}(σ*_j)' should read 'τ_1 = φ^{-1}(σ*_1)'.","section":"Appendix VI-A"},{"comment":"The sentence 'Crucially, these event timestamps depend on the trajectory T(t) taken by the IMU and, by construction, will depend on the event timestamps E*...' presents a consequence of Theorem 1 as a definition; rephrase to avoid appearing circular.","section":"Section III-B"},{"comment":"Events with index j such that j* > B−1 are silently dropped by the Kronecker delta mapping. The paper should state whether M > B occurs on the test data and what the effect is, given that high-acceleration segments may generate more than 200 events per second window.","section":"Section III-D, Eq. (26)"},{"comment":"The first three rows of the table use 100% of the training data, while all other rows use 50%; although the text notes this, the table layout invites direct comparison with the 50% rows. Mark the 100%-data rows more explicitly in the table caption.","section":"Table III"},{"comment":"References [19] and [20] both cite the same RoNIN paper (Herath et al., ICRA 2020) and appear as if distinct; consolidate them to avoid confusion.","section":"References"},{"comment":"The '+ splat.' baseline has catastrophic errors at 20 Hz (ATE* ≈ 192 m), which is a large outlier relative to all other entries. Clarify whether this is an expected artifact of the event-stack interpolation at very low rates or an implementation issue.","section":"Table VI / Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The theoretical result is a straightforward reparametrization property of level-crossing definitions, and the paper does not claim it as more than that, but the proof contains a mismatch with the stated event condition. The more significant concern is the gap between the theorem's factorization assumption and the v0-dependent pre-integration used in practice; this is fixable by either adding a quantitative error analysis or explicitly relegating the theorem to an idealized setting and leaning on the (substantial) empirical evidence. The empirical coverage is a genuine strength, including open-source code and rate-robustness results without rate augmentation. If the authors can tighten the theoretical claims and add basic uncertainty reporting, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuinely useful paper for the inertial odometry crowd, and it deserves a serious referee. The core idea is to replace fixed-rate IMU samples with Lie events—level crossings of the SE(3) pre-integration signal, with polarities normalized to the unit sphere in the Lie algebra. That generalizes event-camera sampling to a manifold, and the invariance claim is that event timestamps and polarities are independent of the time parametrization of the path. If the reference signal factorizes as x(t)=x*(φ(t)), Theorem 1 is correct, and the toy experiment confirms it.\n\nThe paper does a lot of good things: the Lie-group formalism is clean, the polarity normalization is a natural generalization of binary polarities, and the evaluation is unusually broad. They apply the representation to TLIO and RoNIN and test on TLIO, Aria, RIDI, and OxIOD, with rate-sensitivity ablations and hyperparameter studies. The code is open source. The gains are consistent: ATE* drops by 10-20% across datasets at native rates, and the rate-sensitivity curves are much flatter than baselines even without rate augmentation. That last result is the strongest part of the empirical story.\n\nThe soft spot is the gap between Theorem 1 and the actual reference signal. The theorem assumes x(t)=x*(φ(t)) exactly, but the pre-integration in Eqs. (19)-(20) depends on an initial velocity v0. If v0 has error Δv, the pre-integrated position has a term Δv(t-t1) that does not factor as a path reparametrization. So the event timestamps and polarities are 'approximately' invariant, as the paper concedes. The limitations section says pre-integrations suffer from slight drift. This is not fatal: the sensitivity analysis shows graceful degradation under v0 noise, and the whole point is that the representation is robust to that noise. But the theoretical claim is over-stated for the real pipeline, and the paper doesn't provide an error bound or even a precise statement of the approximation. A referee should ask for that, or for a softened claim.\n\nOne more issue, minor relative to the above: all reported numbers are medians without error bars or significance tests. For a systems paper this is common, but the headline 'up to 21%' is a single dataset's MSE, and the EKF-level gains are smaller (3-12%). The paper is honest about this, but the reader should calibrate.\n\nWho is this for? Anyone working on learning-based inertial odometry, and more broadly anyone interested in event-based sampling on manifolds. I'd bring it to our reading group.\n\nRecommendation: send it to peer review. It is a novel representation with substantial empirical support and open code. The theoretical gap should be addressed in revision, but it is not a reason to desk-reject.","headline":"Solid new representation for inertial odometry with a real but fixable gap between its ideal invariance theorem and the actual pre-integration signal.","tokens_in":24816,"tokens_out":3167,"would_cite":true,"duration_ms":32750,"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 that Lie events—IMU samples triggered by level crossings of the SE(3) pre-integration signal—canonicalize inertial inputs against speed and sampling-rate changes, and that training neural displacement priors on these…","keywords":["Lie events","inertial odometry","neural displacement priors","event-based sampling","SE(3)","pre-integration","canonicalization","level-crossing sampling"],"falsifier":"Take one recorded trajectory, resample it into two runs with very different speed profiles, generate Lie events from each, and check whether the polarity vectors agree and whether the event times map onto each other after applying the inverse time warp; if the mismatch is no larger than the injected IMU noise, the canonicalization is confirmed, and if a downstream prior trained on slow motion still fails on fast motion, it is not.","tokens_in":23737,"feed_emoji":"🧭","tokens_out":10003,"duration_ms":93067,"temperature":0.7,"pith_summary":"Neural displacement priors correct inertial odometry drift but generalize poorly to different walking speeds and IMU sampling rates. This paper replaces fixed-rate raw IMU windows with “Lie events”: samples triggered whenever the norm of the change in the $\\mathrm{SE}(3)$ pre-integration crosses a threshold, each carrying a normalized Lie polarity that records the direction of change. A theorem shows that event timestamps, references, and polarities depend only on the underlying path, not on the speed profile, so the network receives canonicalized inputs. Training TLIO and RoNIN displacement priors on event stacks reduces trajectory error on several datasets by up to about 21 percent, with no change to the downstream filter.","feed_headline":"Lie events cut inertial odometry error by up to 21%","feed_subtitle":"Resampling IMU pre-integrations on SE(3) level crossings makes neural priors robust to speed and sampling-rate changes.","key_machinery":"The load-bearing object is the Lie event, defined by level-crossing sampling on $\\mathrm{SE}(3)$. Start from the last reference pose, project the pre-integration onto the tangent space with the Log map, and fire an event when the norm leaves a ball of radius $\\theta$; update the reference to the event pose $x_{\\mathrm{ref},j}=x_{\\mathrm{ref},j-1}\\mathrm{Exp}(\\theta p(\\tau_j))$. Theorem 1 is the identity that carries the argument: because event times are defined through the Log map, a time reparametrization $\\varphi$ merely shifts event times by $\\varphi^{-1}$ and leaves references and polarities unchanged. Event stacks then pack measurements and polarities into a fixed $B\\times 12$ tensor, making the representation directly consumable by off-the-shelf neural displacement priors.","core_discovery":"The paper’s central claim is that event-based sampling of the IMU pre-integration on the Lie group $\\mathrm{SE}(3)$ canonicalizes the input for neural displacement priors. The reference signal $x(t)=(\\tilde R(t),\\tilde t(t))$ is the on-manifold pre-integration of bias-corrected accelerometer and gyroscope readings, and an event fires at the first time $\\tau_j$ when $\\lVert \\mathrm{Log}(x_{\\mathrm{ref},j-1}^{-1}x(\\tau_j))\\rVert=\\theta$; the Lie polarity $p(\\tau_j)=\\mathrm{Log}(x_{\\mathrm{ref},j-1}^{-1}x_{\\mathrm{ref},j})/\\lVert\\mathrm{Log}(x_{\\mathrm{ref},j-1}^{-1}x_{\\mathrm{ref},j})\\rVert\\in S^5$ records the normalized change direction. Theorem 1 states that whenever $x(t)=x^*(\\varphi(t))$ with $\\varphi'>0$, the event times satisfy $\\tau_j=\\varphi^{-1}(\\sigma^*_j)$ and the references satisfy $x_{\\mathrm{ref},j}=x^*_{\\mathrm{ref},j}$, so the polarities are invariant under time reparametrization. Training TLIO and RoNIN on these events reduces absolute translation error on the TLIO, Aria, RoNIN, RIDI, and OxIOD datasets, and the error degrades gracefully when the IMU rate is subsampled from 1 kHz or 800 Hz down to 20 Hz.","pith_inferences":["The theorem is a recipe for any sensor whose output is a time reparametrization of a fixed path: encoders, tactile skins, or depth streams could be event-sampled on their natural manifold and fed to a neural prior, which is the broader conjecture the paper states but does not test.","The residual failure mode is pre-integration drift: since the initial velocity comes from a filter, event timestamps inherit filter uncertainty; replacing the reference signal with a smoother or bias-corrected state estimate would tighten the canonicalization.","A testable extension: train exclusively on slow trajectories and evaluate on fast versions of the same paths; the theory predicts near-equal displacement estimates, which would isolate the role of canonicalization from the network’s learned priors.","The gains from Lie events should stack with yaw-equivariance or test-time-training objectives, since those address orientation variability while Lie events address speed variability; the paper does not combine them."],"forward_implications":["When IMU data is subsampled from 200 Hz to 20 Hz, the event-based TLIO variant keeps most of its accuracy while the interpolation and event-splat baselines degrade sharply, so rate augmentation becomes unnecessary.","The same preprocessing transfers to velocity-regression priors such as RoNIN, reducing absolute translation error on RIDI and OxIOD even when trained on only half the public training data.","Because the representation canonicalizes speed, the largest gains appear on out-of-distribution high-speed motion, exactly where raw-IMU priors fail.","The lower event rate (about 74 Hz versus 200 Hz of raw IMU data) means the approach also removes input redundancy without sacrificing accuracy."],"supporting_citations":[{"why":"Supplies the TLIO neural displacement prior, its EKF measurement model, and the 60-hour IMU dataset used for training and evaluation.","marker":"[31]"},{"why":"Supplies the RoNIN velocity-regression prior and its pedestrian dataset, plus the RIDI and OxIOD evaluation protocol.","marker":"[20]"},{"why":"Provides the on-manifold pre-integration equations that define the reference signal x(t) in SE(3).","marker":"[13]"},{"why":"Supplies the event-based vision paradigm whose level-crossing sampling and event stacks are generalized to Lie groups.","marker":"[15]"},{"why":"Defines the original dynamic vision sensor whose binary polarity and threshold event model the paper generalizes.","marker":"[29]"},{"why":"Contributes the event-triggered attitude-control formulation on SO(3) that the paper extends to SE(3) with polarities.","marker":"[24]"},{"why":"Provides the MSCKF ground-truth trajectories used to supervise and evaluate TLIO-trained models.","marker":"[36]"},{"why":"Supplies the Aria Everyday Activities dataset used as the out-of-distribution generalization test across IMU rates and activities.","marker":"[33]"}],"fun_headline_variants":["Lie events cut IMU odometry error up to 21%","Event-based IMU sampling trims odometry drift by 21%","Robust inertial odometry via Lie-event priors","Lie events slash inertial odometry error 21%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the IMU pre-integration is exactly a time reparametrization of a fixed path; in practice it is computed from noisy, biased readings with an initial velocity supplied by a filter, so the path-only canonicalization holds only approximately.","fun_headline_variants_meta":{"raw":{"variants":["Lie events cut IMU odometry error up to 21%","Event-based IMU sampling trims odometry drift by 21%","Robust inertial odometry via Lie-event priors","Lie events slash inertial odometry error 21%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000379,"raw_usage":{"total_tokens":2090,"prompt_tokens":1096,"completion_tokens":994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":924}},"tokens_in":712,"tokens_out":994,"duration_ms":8018,"temperature":1.0,"reasoning_tokens":924,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:24:11.827947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one recorded trajectory, resample it into two runs with very different speed profiles, generate Lie events from each, and check whether the polarity vectors agree and whether the event times map onto each other after applying the inverse time warp; if the mismatch is no larger than the injected IMU noise, the canonicalization is confirmed, and if a downstream prior trained on slow motion still fails on fast motion, it is not.","supporting_citations":[{"cited_title":"Ronin: Robust neural inertial navigation in the wild: Benchmark, evaluations, and new methods","cited_arxiv_id":null,"evidence_quote":"Supplies the RoNIN velocity-regression prior and its pedestrian dataset, plus the RIDI and OxIOD evaluation protocol."},{"cited_title":"Mourikis and Stergios I","cited_arxiv_id":null,"evidence_quote":"Provides the MSCKF ground-truth trajectories used to supervise and evaluate TLIO-trained models."}],"review_version":1}