{"id":"3e6044c4-492c-43f6-9029-655d2f4000a8","arxiv_id":"2502.05491","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Lie-algebra error-state linearization plus least-squares parameter identification and LQR is proposed for adaptive rigid-body trajectory tracking under mass and inertia uncertainty.","lead":"This paper proposes a control method that lets a rigid-body robot track a moving reference trajectory even when its mass and inertia are not exactly known. It converts the robot's motion into a simpler linear error model, estimates the unknown parameters from collected data, and shows lower tracking errors than two vector-space adaptive baselines in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No validity region is given for the linearization in Proposition 1; if exploration noise or initial error leaves that region, the least-squares reconstruction (19) is biased and the DARE gain is based on the wrong (A,B).","rationale":"The reader's conditional verdict is appropriate: the core idea is coherent and the simulations are encouraging, but the linearization in Proposition 1 is doing double duty — it is both the control-synthesis model and the data-generating model for parameter identification. The weakest point is therefore the unquantified validity region of that linearization. I agree with the reader's weakest-assumption identification. The concern is load-bearing because if the exploration noise or initial error pushes the state outside the linear regime, the least-squares problem in Eq. (19) is fitting the wrong model, and the recovered inertia and mass are biased even with infinite data. The paper supplies no error bound, no restriction on γ, no consistency theorem, and no stability analysis for the adaptive loop. The ridge regularization in Eq. (19) introduces a further bias that is not analyzed. The available source code is a real strength and makes the proposed test straightforward to run. This concern does not move the reader's verdict: the paper should remain conditional, with the required revision being either a validity-region/error-bound analysis or a demonstration that the reconstruction is insensitive to exploration-noise amplitude and initial error. If the proposed test shows large bias, the central identification claim would be substantially weakened.","tokens_in":8875,"tokens_out":15729,"duration_ms":144785,"concrete_test":"Use the authors' public code to generate two datasets with identical exploration-noise realization and initial condition: (i) sample the true continuous-time dynamics (4) at Δt, and (ii) sample the discrete linear model (15) with ground-truth J_b. Apply the same ridge least-squares reconstruction (19) to both and compare reconstructed (Î_b, m̂) against ground truth as N grows and as the standard deviation of γ is varied. If the nonlinear-data reconstruction error does not approach the linear-data error level (or zero) and instead plateaus at a value that grows with noise variance or initial pose error, the linearization assumption is load-bearing. A secondary check: halve Δt; if the reconstruction error changes materially, the forward-Euler discretization in Eq. (15) is also a source of bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1 derives Eq. (9) by replacing exp(ψ^) with I + ψ^ and by first-order linearizing the twist dynamics about {ζ_d, u_d}, and Eq. (15) turns this into the discrete linear model used in Eq. (18). Algorithm 1, however, collects its data from the true nonlinear dynamics (4) under an exploration input u = u_d + Kx + γ. Nothing in the paper bounds the remainder terms O(‖ψ‖^2, ‖δζ‖^2) or constrains the amplitude of γ or the size of the initial error, so the regressor [X, U] in Eq. (19) is not guaranteed to be generated by the linear model. The least-squares estimate of (A,B) is then a biased fit of a nonlinear map; because B = [0; J_b^{-1}]Δt, any bias in B directly biases the reconstructed inertia and mass, and the DARE gain computed from the estimated (A,B) inherits that bias. The ridge term λ in Eq. (19) adds an additional, unquantified bias. The paper provides no convergence or consistency proof for the reconstruction and no closed-loop stability analysis, so the central claim that model parameters are reconstructed from data is only supported by the particular simulation setup. This is the load-bearing assumption because it is exactly the step that turns nonlinear rigid-body data into a linear-identification problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Lie-algebra-based adaptive tracking control method for rigid body dynamics on SE(3). It transforms the pose error into the Lie algebra, linearizes the error dynamics to obtain a linear state-space model (Eq. (9) and its discretization Eq. (15)), and then uses regularized least squares on collected state/input data to reconstruct the unknown mass and inertia matrix. The reconstructed parameters are used to solve a DARE for an LQR feedback policy. The paper reports Monte Carlo simulations showing parameter reconstruction errors decreasing with dataset size and substantially smaller tracking errors than OFU and TS baselines. The source code is publicly available.","tokens_in":9262,"tokens_out":4428,"duration_ms":44055,"significance":"The core idea of decoupling model parameters from the system state via the Lie-algebra representation is appealing and could offer computational and geometric advantages over vector-space adaptive methods. The exact error dynamics in Lemma 1 are correctly derived, and the identification formulation is coherent. The paper also provides reproducible code, which strengthens its practical value. However, the central claim that model parameters can be reliably reconstructed from data rests on an unquantified linearization and an unproven identification step. If these gaps are addressed, the work could be a useful contribution; in its current form, the theoretical support is incomplete.","major_comments":[{"comment":"The derivation of the linear error dynamics relies on the first-order Taylor approximation exp(ψ^) ≈ I + ψ^ and a first-order linearization of the twist dynamics about {ζ_d, u_d}, but no validity region or error bound is provided. The data used in Algorithm 1 are collected from the true nonlinear dynamics (4) under exploration noise γ, and nothing in the paper bounds the remainder terms O(‖ψ‖², ‖δζ‖²) or constrains the amplitude of γ or the initial tracking error. Consequently, the regressor in Eq. (18)–(19) is not guaranteed to be generated by the linear model, and the least-squares estimate of (A, B) may be biased. Because B = [0; J_b^{-1}]Δt, any bias in B directly biases the reconstructed inertia and mass, and the DARE gain inherits that bias. The paper should either provide a rigorous error bound with explicit conditions on the exploration and initial state, or weaken the claims accordingly.","section":"Section III-A, Proposition 1, Eq. (9)"},{"comment":"Algorithm 1 has no stability, convergence, or consistency theorem. The paper does not prove that the regularized least-squares solution of Eq. (19) converges to the true (A, B) as N increases, nor that the resulting DARE-based feedback stabilizes the true nonlinear closed-loop system. The only support for the central claim that model parameters are reconstructed from data is the particular simulation setup. A formal statement of the identification error, conditions for consistency (e.g., persistence of excitation and linearization validity), and a closed-loop stability guarantee are needed. Without these, the method is best described as a heuristic data-driven LQR, rather than a certified adaptive controller.","section":"Section III-B, Algorithm 1"},{"comment":"The paper does not specify how the least-squares estimate of B is converted into a physically meaningful inertia matrix and mass. The true B has the block structure B = [0; J_b^{-1}]Δt, but the unconstrained estimate from Eq. (19) will generally not have exactly this structure, and the reconstructed bottom block may not be symmetric or positive definite. The paper merely states \"Reconstruct (I_b, m) by (15)\" without describing the extraction, symmetrization, or projection procedure, or how any non-idealities in the estimate affect the reconstructed parameters. This step is load-bearing for the claimed parameter reconstruction and should be made explicit.","section":"Section III-B, Eq. (15) and reconstruction step"}],"minor_comments":[{"comment":"The paper describes the Lie-algebra transformation as yielding a \"globally\" valid vector-space representation, but the linearization in Proposition 1 is local. The global claims in the abstract and introduction should be softened to avoid overstating the validity region.","section":"Section II-A and Section III-A"},{"comment":"The algorithm's pseudocode refers to an initial gain K and uses u_k = K x_k + u_d + γ_k, but K is never defined or initialized in the algorithm block. Clarify how K is selected before the data collection phase.","section":"Algorithm 1"},{"comment":"The text says \"regulation term\" but should be \"regularization term\". Also, the dimension of the identity matrix in λI is not stated; specifying it would improve clarity.","section":"Section III-B, Eq. (19)"},{"comment":"Tracking error comparisons in Table I are reported as single numbers without error bars, confidence intervals, or information about the number of Monte Carlo runs shown. Since the evaluation is central to the claimed superiority, reporting mean±standard deviation or interquartile ranges would strengthen the comparison.","section":"Section IV, Table I"},{"comment":"The caption says the figure depicts the \"evolution\" of reconstruction errors, but it is unclear whether these curves are Monte Carlo means, medians, or individual runs. Please state the statistic and add error bars.","section":"Section IV, Figure 2"},{"comment":"The reference pose is said to be obtained by integrating from \"the initial pose at the origin, i.e., X0 = I\"; the identity matrix corresponds to the origin in position with identity orientation, but the phrasing is ambiguous. Rephrase for clarity.","section":"Section IV, simulation setup"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a promising geometric idea with reproducible code, but the current theoretical gaps—particularly the unquantified linearization and the absence of any identification or stability guarantees—prevent acceptance in its present form. The authors should be encouraged to add either rigorous analysis or a clear statement of the method's heuristic nature, and to describe the parameter extraction step in detail. The comparison with OFU and TS would also benefit from statistical reporting. Overall, the paper is within the scope of the journal but needs substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a competent, modest paper. It takes the known error-state linearization on SE(3) (Teng et al., the GMPC line) and combines it with the well-worn regularized least-squares + DARE adaptive control recipe from vector-space LQR. The new bit is the specific pairing — identify mass and inertia from the Lie-algebra error dynamics rather than from a Euclidean model. That is a legitimate incremental contribution, and the geometry is handled correctly. Lemma 1 is exact, and Proposition 1 is a standard first-order linearization, clearly derived. The paper gives credit where prior work did the heavy lifting, and the code is public, which matters.\n\nThe simulations are decent. The reconstruction errors do drop as the dataset grows, and the improvement over OFU and Thompson Sampling looks real — not huge in absolute terms, but the proposed method's errors are one to two orders smaller across the board. My main gripes are the ones you flagged. There is no validity region for the Proposition 1 linearization. The algorithm collects data from the true nonlinear dynamics under exploration noise, then fits a linear model. Nothing bounds the remainder terms or the exploration amplitude, so the least-squares reconstruction can be biased for states that stray far from the operating point. The ridge term adds another unquantified bias. I also miss any statement about convergence or closed-loop stability of the adaptive loop; as written, the method is supported only by these particular simulations. That is a genuine gap for a control paper, not a nitpick.\n\nTable I is another soft spot. No error bars, no statement on how many Monte Carlo runs, and no details on how the baselines were tuned. The headline comparison is under-reported, even though the qualitative story is probably right.\n\nThat said, I do not think the central idea is broken. The linearization concern is real but likely manageable in practice if the exploration is small and the controller keeps the error close to zero. The authors should be asked to quantify the region, give a convergence or consistency argument under the linearized model, and report statistics on the comparison. This is a publishable workshop or conference paper with revision, and I would not desk-reject it. The right call is to send it to a serious referee who can press on the identification bias and the missing stability analysis.","headline":"A clean but incremental Lie-algebra parameter-identification scheme for SE(3) tracking, with honest simulations and public code; the missing linearization validity region and lack of convergence guarantees are the real soft spots.","tokens_in":9698,"tokens_out":887,"would_cite":false,"duration_ms":11264,"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":"By lifting the SE(3) pose error into its Lie algebra and linearizing about the reference trajectory, this paper derives a linear error model whose unknown mass and inertia appear only in the state and input matrices, so regularized least…","keywords":["adaptive control","Lie-algebraic control","rigid body dynamics","trajectory tracking","SE(3)","optimal control","parameter identification","linear quadratic regulator"],"falsifier":"Run Algorithm 1 with an initial pose error far from the operating point (large $\\|\\psi(0)\\|$ and $\\|\\delta\\zeta(0)\\|$) and with large exploration noise, then compare the next-state prediction of the identified linear model Eq. (15) against the true simulated next state; if the prediction error does not shrink with dataset size, or if the reconstructed $m$ and $I_b$ stay biased, the decoupling claim fails outside the small-error regime.","tokens_in":8703,"feed_emoji":"🎯","tokens_out":11294,"duration_ms":95312,"temperature":0.7,"pith_summary":"The paper proposes an adaptive tracking controller for a fully actuated rigid body whose mass and inertia are unknown. Its central move is to map the $SE(3)$ pose error into the Lie algebra $\\mathfrak{se}(3)$ and take a first-order Taylor approximation, turning the nonlinear manifold error dynamics into the linear system of Eq. (9) and its discrete version Eq. (15). In that linear model, the uncertain parameters $m$ and $I_b$ appear only in the matrices $A_{\\zeta_d}$ and $B$, decoupled from the current error state. Algorithm 1 uses a regularized least-squares fit on collected state and input data to recover those parameters, then solves a discrete algebraic Riccati equation for the optimal feedback gain. In Monte Carlo simulations the reconstructed inertia and mass converge as the dataset grows, and the resulting tracking errors are roughly one to two orders of magnitude smaller than those of the OFU and TS baselines.","feed_headline":"Lie-algebra adaptive control cuts rigid-body tracking error up to 69x","feed_subtitle":"Mapping SE(3) errors to the Lie algebra lets one least-squares fit identify mass and inertia for an optimal tracker.","key_machinery":"The central mechanism is the vector-space image of the $SE(3)$ error dynamics under the exponential and logarithm maps. The paper uses the first-order expansion $\\exp(\\psi^{\\wedge}) \\approx I + \\psi^{\\wedge}$ to convert the pose error equation into $\\dot{\\psi} = -\\mathrm{ad}_{\\zeta_d}\\,\\psi + \\delta\\zeta$, and linearizes the twist equation about $\\{\\zeta_d, u_d\\}$ to obtain $\\Gamma_{\\zeta_d}$. The resulting pair $(A_{\\zeta_d}, B)$ carries all model parameters while remaining error-state independent, which is what makes parameter reconstruction a single least-squares problem and control synthesis a single DARE solve.","core_discovery":"Proposition 1 is the load-bearing result. Defining the pose error in the Lie algebra as $\\psi(t)^{\\wedge} = \\log(X_d(t)^{-1} X(t))$ and the perturbed input as $\\delta u = u - u_d$, the paper obtains the linear error dynamics $\\dot{\\psi} = -\\mathrm{ad}_{\\zeta_d}\\,\\psi + \\delta\\zeta$ and $\\dot{\\delta\\zeta} = \\Gamma_{\\zeta_d}\\,\\delta\\zeta + J_b^{-1}\\,\\delta u$, i.e. Eq. (9). Because the state transition matrix depends only on the reference trajectory and the model parameters, and not on the current error state, the parameters $m$ and $I_b$ can be read back out of the identified $(A_{\\zeta_d}, B)$ by solving the regularized least-squares problem Eq. (19). Solving the DARE with the reconstructed model then yields the optimal feedback policy for Problem 1.","pith_inferences":["The same decoupling should transfer to other matrix Lie groups used in robotics, such as $SE(2)$ or the connected Lie groups used for legged-robot state estimation, since the derivation only uses the adjoint map and the exponential map; the paper demonstrates only $SE(3)$.","A recursive least-squares version of Eq. (19) with forgetting would let the estimates track slow parameter drift online, which the batch formulation in Algorithm 1 does not address.","The exploration noise is justified by a persistence-of-excitation condition, but no quantitative richness measure is given; one could test directly whether the condition number of the regressor matrix in Eq. (19) predicts the observed reconstruction error.","If the linear model remains valid, the same dataset could be reused to estimate external disturbances or actuator faults by stacking them into the unknown vector, an extension the paper does not explore."],"forward_implications":["Once $m$ and $I_b$ are reconstructed, the optimal feedback gain follows from one DARE solve, so the method runs in real time; the paper reports an average computation time of about 0.17 s for $N = 2000$ data points.","The identified linear model can be reused in a time-varying LQR or MPC formulation for non-stationary reference trajectories, because the quadratic objective with linear constraints becomes a standard quadratic program.","Because the error state lives in the vector space $\\mathfrak{se}(3)$, the approach avoids local coordinate singularities such as gimbal lock while retaining the geometric consistency of the group description.","In the paper's simulations, the adaptive controller reduces position, rotation, angular-velocity, and linear-velocity tracking errors by one to two orders of magnitude relative to the OFU and TS LQR baselines."],"supporting_citations":[{"why":"Supplies the exponential/logarithm map and the adjoint operation used to lift the SE(3) error into the Lie algebra.","marker":"[8]"},{"why":"Provides the rigid-body dynamics model with generalized inertia J_b that the paper linearizes.","marker":"[7]"},{"why":"Motivates designing the control objective in the Lie algebra and is the geometric basis for the error-state formulation.","marker":"[11]"},{"why":"Gives the standard discretization of the continuous-time linear system into Eq. (15) and the DARE used for the optimal gain.","marker":"[20]"},{"why":"Defines the OFU-LQR baseline and the regularized least-squares parameter-reconstruction style the paper adapts.","marker":"[15]"},{"why":"Defines the TS-LQR baseline and contributes the same regularization scheme.","marker":"[22]"},{"why":"Supplies the persistence-of-excitation condition that justifies the exploration noise in Algorithm 1.","marker":"[23]"},{"why":"Gives the system-identification perspective that motivates the regularization term in Eq. (19).","marker":"[21]"}],"fun_headline_variants":["Lie-algebra trick cuts rigid-body tracking error 69x","Adaptive control on Lie algebra yields 69x error drop","Rigid-body tracking: Lie-algebra method slashes error 69x","SE(3) log-map enables optimal adaptive tracking, 69x less error","Optimal adaptive control via Lie algebra: 69x error reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the first-order Taylor approximation $\\exp(\\psi^{\\wedge}) \\approx I + \\psi^{\\wedge}$, together with the linearization of the twist dynamics about $\\{\\zeta_d, u_d\\}$, remains accurate for every state and input the exploration visits; the paper gives no error bound or validity region for this approximation.","fun_headline_variants_meta":{"raw":{"variants":["Lie-algebra trick cuts rigid-body tracking error 69x","Adaptive control on Lie algebra yields 69x error drop","Rigid-body tracking: Lie-algebra method slashes error 69x","SE(3) log-map enables optimal adaptive tracking, 69x less error","Optimal adaptive control via Lie algebra: 69x error reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1203,"prompt_tokens":882,"completion_tokens":321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":226}},"tokens_in":498,"tokens_out":321,"duration_ms":3569,"temperature":1.0,"reasoning_tokens":226,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:06:19.934209+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 with an initial pose error far from the operating point (large $\\|\\psi(0)\\|$ and $\\|\\delta\\zeta(0)\\|$) and with large exploration noise, then compare the next-state prediction of the identified linear model Eq. (15) against the true simulated next state; if the prediction error does not shrink with dataset size, or if the reconstructed $m$ and $I_b$ stay biased, the decoupling claim fails outside the small-error regime.","supporting_citations":[{"cited_title":"Lie algebraic cost function design for control on Lie groups,","cited_arxiv_id":null,"evidence_quote":"Motivates designing the control objective in the Lie algebra and is the geometric basis for the error-state formulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the standard discretization of the continuous-time linear system into Eq. (15) and the DARE used for the optimal gain."},{"cited_title":"Regret bounds for robust adaptive control of the linear quadratic regulator,","cited_arxiv_id":null,"evidence_quote":"Defines the OFU-LQR baseline and the regularized least-squares parameter-reconstruction style the paper adapts."},{"cited_title":"Thompson sampling for linear-quadratic control problems,","cited_arxiv_id":null,"evidence_quote":"Defines the TS-LQR baseline and contributes the same regularization scheme."},{"cited_title":"Perspectives on system identification,","cited_arxiv_id":null,"evidence_quote":"Gives the system-identification perspective that motivates the regularization term in Eq. (19)."}],"review_version":1}