{"id":"33fe5401-2cb4-4dde-b433-64c4b635031e","arxiv_id":"2504.20009","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"STELA combines a sampling-based kinodynamic plan with factor-graph optimization to estimate the past trajectory and adapt future controls in real time for nonlinear robot models.","lead":"STELA is a factor-graph method that simultaneously estimates a robot's past trajectory and adapts its future controls to follow a planned path under noise and model errors. It extends earlier factor-graph control work from linear holonomic models to nonlinear vehicles, and it is tested in simulation and on a real MuSHR car.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generality to 'any dynamical system' is not established: the dynamics factor of Section V-B models q̈=f(u) only, so state-dependent accelerations (gravity, drag, Coriolis) are outside the stated factor equations; a nonlinear second-order test would settle it.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: Section V-B restricts the dynamics factor to control-only acceleration, so the abstract's 'any dynamical system' claim is not supported by the equations. This is the primary advertised difference from STEAP and SCATE, which the paper says are limited to holonomic, linear-time-varying models. If the factor graph cannot represent state-dependent dynamics, the novelty claim shrinks to a useful but narrower class: control-driven acceleration models with Lie-group integration and variable edge durations. The experiments are consistent with that narrower class, and the real-robot results give genuine support for the practical method. No internal contradiction was found in the tested regime, and the Limitations section is honest about extreme-noise failures. The lack of code and confidence intervals is secondary to the model-class overclaim. Because the reader's verdict was already CONDITIONAL and this concern confirms rather than redirects it, the appropriate output is UNCHANGED.","tokens_in":18866,"tokens_out":4944,"duration_ms":54046,"concrete_test":"Implement STELA with exactly the Section V-B dynamics factor for a damped pendulum or cart-pole, q̈ = -(g/l)sin(q) - (b/ml²)q̇ + u/(ml²), using the same SBMP initialization and sliding-window procedure. Run closed-loop trajectory following from several initial states and compare against ground-truth simulation of the true state-dependent dynamics. If the factor graph can track only after adding state-dependent terms to the dynamics factor, the general-form claim fails. A cheaper analytical check is to write out the predicted q̇_{i+1}=q̇_i+f(u_i)Δt for this system; because the true update depends on q_i and q̇_i, the equalities cannot hold for arbitrary states, independent of tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central advertised contribution is that STELA's factor graph 'can be applied for any dynamical system given access to first or second-order state update equations.' The construction in Section V-B undercuts this. The dynamics factor defines acceleration solely as a system-specific function of the control input, q̈=f(u_i), and then predicts q̇_{i+1}=q̇_i+q̈Δt_i via Euler integration. There is no term f(q,q̇,u). The integration factor q_{i+1}=q_i∘Exp(q̇_iΔt_i) updates configuration using the initial velocity only, omitting the 0.5·q̈·Δt² correction that a second-order state update would contribute. Thus the graph solves a restricted model: piecewise-constant controls with linear velocity extrapolation, independent of configuration- or velocity-dependent forces. For a true first-order system ẋ=f(x,u), the graph treats q̇ as an auxiliary velocity variable rather than encoding f(x,u); no first-order factor is defined. The two experimental models fit the restricted form: LTV-SDE has acceleration equal to control, and the MuSHR model uses control-driven acceleration and steering angle. Neither exercises gravity, drag, Coriolis, friction maps, or other state-dependent accelerations, so the 'any dynamical system' claim is an unsupported extrapolation rather than an internal inconsistency of the method on the tested class. The Limitations section acknowledges failure under extreme noise but does not qualify the model-class claim, which makes the overclaim part of the paper's central thesis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STELA, a factor-graph framework for simultaneous trajectory estimation and local control adaptation during kinodynamic trajectory following. A kinodynamic sampling-based motion planner first generates a feasible, collision-free trajectory, which is converted into a factor graph. STELA then runs sliding-window incremental inference with iSAM2 to (i) estimate past robot states from noisy observations and (ii) locally adapt future controls, with the duration of each trajectory edge treated as an optimization variable. The method is evaluated in simulation on an LTV-SDE holonomic model and on a second-order MuSHR car model, compared against open-loop execution, SCATE variants, and SBMP replanning, and demonstrated on a real MuSHR robot. Ablation studies examine the sliding window, duration variables, obstacle factors, and SBMP initialization.","tokens_in":19179,"tokens_out":5653,"duration_ms":59734,"significance":"If its claims are appropriately scoped, STELA is a solid engineering contribution to integrated estimation and control. It demonstrates, for the first time in this factor-graph line of work, successful online (10 Hz and above) simultaneous trajectory estimation and control on a non-holonomic, second-order vehicle, and it adds variable edge duration as an optimization variable. The experimental section is comparatively thorough, with multiple baselines, ablations, and real-robot validation under a substantial model gap, and the paper clearly identifies the SBMP plan as the source of the prior. The central 'any dynamical system' claim, however, is not supported by the factor equations, and the success-rate comparisons lack confidence intervals and are based on small trial counts in several scenes; both issues need to be addressed before the paper's advertised contributions can be taken at face value.","major_comments":[{"comment":"The paper's advertised generality to 'any dynamical system given access to first or second-order state update equations' is not supported by the factor definitions. The dynamics factor constrains acceleration via a system-specific function q̈=f(u_i), with no dependence on q_i or q̇_i, and the integration factor predicts q_{i+1}=q_i∘Exp(q̇_i Δt_i), omitting the 0.5·q̈·Δt_i² term of a second-order update. Consequently, state-dependent forces such as gravity, drag, Coriolis effects, or friction cannot be encoded, and no first-order state-update factor ẋ=f(x,u) is defined. The two test models (LTV-SDE and MuSHR) both fit the restricted control-only-acceleration form, and the Limitations section does not qualify the model-class claim. Please either replace the 'any dynamical system' claim with a precise characterization of the supported model class, or extend the factor graph with state-dependent dynamics factors and validate on a nonlinear second-order system with state-dependent acceleration.","section":"V-B (dynamics and integration factors)"},{"comment":"The central success-rate comparisons are reported as point estimates without confidence intervals or significance tests, and several cells rest on very few trials: Table V gives one start-goal trajectory with five repetitions for Simple Obstacle and two trajectories with five repetitions for Bug Trap. Claims such as 'STELA achieves at least comparable performance' and 'maintains a high success rate' should be supported by binomial confidence intervals (or equivalent) and, where differences are modest, by a comparison against SCATE-SBMP; otherwise the reader cannot assess whether differences are within sampling noise.","section":"VI-A, Tables I-VII and Table V"},{"comment":"The statement that STELA 'does not make any assumptions in terms of Gaussian priors for the underlying processes' is contradicted by the construction: every factor has the form exp(−0.5·||h||²_Σ), and the SBMP initialization is injected through unary prior factors on q, q̇, and Δt with covariance matrices. The novelty is better described as using a feasible kinodynamic plan as the prior rather than hand-tuned trajectory priors. Please rephrase the claim to remove this inconsistency.","section":"V-A and IV (prior and Gaussian assumptions)"}],"minor_comments":[{"comment":"The caption says 'LVT-SDE' but the model is LTV-SDE; please correct the typo.","section":"Table I caption"},{"comment":"In the paragraph introducing the obstacle factor, 'Th factor' should read 'The factor.'","section":"V-B (obstacle factor)"},{"comment":"'Indeterminant Linear System Exception' should be 'Indeterminate Linear System Exception.'","section":"VIII (Limitations)"},{"comment":"The MuSHR dynamics model is described only verbally; writing out the dynamics equations (or reference) would make the restricted model class and the controlled-acceleration assumption transparent.","section":"VI-A (MuSHR model)"},{"comment":"The expression p(θ_STELA|e) ∝ f_TE f_LA should explicitly include the product or conjunction of the two factor groups; as written it looks like a product of two functions without a defined operation.","section":"V-C (Eq. 1)"}],"recommendation":"major_revision","confidential_remarks":"The 'any dynamical system' claim appears in the abstract and introduction and is therefore not a minor wording issue; it is central to the paper's positioning. The engineering results are otherwise credible and worth publishing after the model-class claim is re-scoped and the statistical support for the headline success rates is strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental robotics paper. The genuinely new piece is the combination of SBMP-generated initialization, duration as an optimization variable, and sliding-window iSAM2 updates for simultaneous trajectory estimation and local adaptation. On the LTV-SDE and MuSHR models tested, it works and beats sensible baselines, including real-robot runs. Worth engaging.\n\nWhat it does well: the factor graph construction is clear, the Lie-group integration factor is sensible, and the sliding window keeps computation at 20–30 Hz. The ablation study actually isolates the components—window size, duration variable, obstacle factor, and SBMP initialization—which is more than many papers in this area do. The real MuSHR experiments, including the ramp and movable obstacles, are genuine evidence. Reimplementing SCATE in C++ for a fair comparison is also good practice.\n\nThe main problem is the generality claim. The dynamics factor only models acceleration as a function of control, q̈ = f(u_i), and predicts velocity via Euler integration. There is no term for configuration- or velocity-dependent forces. For a general second-order system ẍ = f(x, ẋ, u), this factor form is wrong. The tested models fit the restricted form—LTV-SDE has acceleration equal to control, and the MuSHR model uses control-driven acceleration and steering angle—so the method is fine on those, but the abstract and Section V-B overstate applicability to 'any dynamical system.' A nonlinear test with state-dependent acceleration, like a pendulum or a quadrotor with drag, would settle it.\n\nSecondary issues: success rates are reported without confidence intervals, and some cells come from only five repetitions, so differences like 0.72 vs 0.88 are not meaningful. No code or full parameter settings are released, which hurts reproducibility. The Limitations section is honest about failure under extreme noise, but it does not qualify the model-class claim, so the overclaim remains part of the central thesis.\n\nBottom line: the method is plausible and the experiments support it on the tested class. The overclaim should be fixed before publication. I would send this to peer review—the core contribution is real, the evaluation is decent, and the main flaw is fixable with a more careful claim and ideally an additional nonlinear test.","headline":"Good incremental extension of factor-graph control with a real-robot evaluation, but the 'any dynamical system' claim is not supported by the dynamics factor actually implemented.","tokens_in":19712,"tokens_out":1704,"would_cite":true,"duration_ms":17938,"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 single factor graph simultaneously estimates a robot's past trajectory and adapts its controls, for any dynamical system with first- or second-order state updates, starting from a kinodynamic sampling-based plan.","keywords":["factor graphs","simultaneous trajectory estimation","local adaptation","kinodynamic motion planning","non-holonomic robots","sliding window inference","incremental smoothing","sampling-based motion planning"],"falsifier":"Simulate a damped double integrator with known damping coefficient $c$, so the true dynamics are $­\\ddot{q} = -c \\dot{q} + u$, and run STELA with the paper's dynamics factor, which predicts $­\\dot{q}_{i+1} = \\dot{q}_i + f(u_i)\\Delta t$ without the $-c \\dot{q}$ term. If the forward predictions become systematically biased and the success rate falls as $c$ grows, the claim that the representation applies to any dynamical system with first- or second-order state updates is not supported beyond the control-affine form.","tokens_in":18636,"feed_emoji":"🤖","tokens_out":10135,"duration_ms":101422,"temperature":0.7,"pith_summary":"STELA asks whether one probabilistic factor graph can do both jobs that autonomous robots usually split: estimate where the robot has been and decide what controls to send next. Earlier factor-graph methods for this combined task worked only for idealized holonomic robots with linear time-varying models and depended heavily on hand-chosen Gaussian priors. STELA claims to remove those restrictions by initializing with a feasible trajectory from a kinodynamic sampling-based planner and then solving one optimization that simultaneously smooths the past trajectory and adapts future controls, with the time between states as an optimization variable. If the claim holds, a robot with non-linear, second-order dynamics and a large gap between its planning model and reality can still follow a collision-free trajectory online at control rates around 10–30 Hz.","feed_headline":"One factor graph both tracks and steers a robot in one pass","feed_subtitle":"The STELA method handles car-like, non-linear robots online, with control updates at 10–30 Hz.","key_machinery":"The central object is the STELA factor graph and the maximum a posteriori inference performed over it. In this context, a factor graph is a bipartite graph in which variable nodes hold robot states, controls, and edge durations, while factor nodes penalize violations of dynamics, observations, priors, obstacle clearance, and control or duration limits. What carries the argument is the factorization of the posterior into a trajectory-estimation product over a short history and a local-adaptation product over a forward horizon, both re-optimized incrementally with a sliding window so the graph size stays bounded as new observations arrive. The duration variable is the distinctive mechanism: it lets the optimizer stretch or shrink the time allocated to each edge, which is how the method absorbs execution noise and recovers from model error without abandoning the planned, collision-free path.","core_discovery":"The paper's central claim is that simultaneous trajectory estimation and local adaptation can be written as inference in one factor graph, with posterior proportional to the product of a trajectory-estimation term and a local-adaptation term. The graph is initialized node-by-node from the edge representation of a kinodynamic sampling-based plan, so the initial guess is dynamically feasible and collision-free rather than a naive straight line. Each edge carries configuration, velocity, control, and duration variables; the duration of execution between two states is optimized, and the integration factor uses forward integration on the configuration manifold. The paper argues that this representation applies to any dynamical system expressible by first- or second-order state update equations, and it demonstrates the method on an idealized linear time-varying stochastic differential equation model and on a second-order car-like robot, including real-robot runs with control updates at 10–30 Hz. Across the reported experiments, STELA matches or exceeds previous factor-graph control methods on the idealized model and succeeds on the car-like model where those methods do not directly apply.","pith_inferences":["Beyond the paper, the “any dynamical system” claim should be read as bounded by the dynamics factor's form: the paper's acceleration model is $­\\ddot{q} = f(u_i)$, so systems with acceleration that depends on velocity or pose, such as damping, friction, or slip, would need an extended factor rather than the exact same graph.","Beyond the paper, the duration variable suggests a concrete test: measure whether STELA's collision avoidance under noise comes mostly from timing adaptation rather than from reshaping the path; if so, combining STELA with a path-replanning module would extend it to dynamic obstacles and to new homotopy classes, which the paper itself notes as future work.","Beyond the paper, the sliding window size is a tuning knob with a predicted trade-off: a longer history gives smoothing benefits and a longer forward horizon gives anticipatory obstacle avoidance, while both increase computation; the ablation results support this, so an adaptive window-size policy is a plausible extension."],"forward_implications":["For robots whose acceleration can be written as a function of control input, the same factor graph now covers both estimation and control, so a separate low-level controller and state estimator do not have to be engineered for trajectory following.","Using a feasible kinodynamic plan as initialization removes the naive straight-line prior that made earlier factor-graph planners fall into local minima, so obstacle-rich environments become more tractable.","Optimizing edge durations lets the controller slow down or speed up automatically, and the reported experiments show this timing adaptation is what keeps trajectories collision-free under high noise.","The sliding window plus incremental updates keeps the factor graph bounded, making online control updates at 10–30 Hz practical rather than requiring full re-optimization of a long trajectory.","Because the graph smooths a short history instead of filtering only the latest state, it can correct past state estimates while choosing the next control, which supports the reported robustness to observation noise."],"supporting_citations":[{"why":"Supplies the factor graph formalism and MAP inference background on which STELA's representation builds.","marker":"[7]"},{"why":"Provides the iSAM2 incremental smoothing and mapping algorithm used for efficient factor graph updates.","marker":"[18]"},{"why":"Provides the earlier iSAM incremental smoothing algorithm that motivates STELA's high-frequency updates.","marker":"[19]"},{"why":"The closest prior method, SCATE, which STELA extends and compares against; limited to linear time-varying dynamics and constant duration steps.","marker":"[22]"},{"why":"The asymptotically optimal kinodynamic sampling-based planner whose output initializes the STELA factor graph.","marker":"[23]"},{"why":"Prior simultaneous trajectory estimation and planning framework that STELA generalizes beyond Gaussian-process dynamics.","marker":"[29]"},{"why":"Source of the obstacle factor definition and of the continuous-time Gaussian process approach that previously limited such methods to holonomic linear systems.","marker":"[30]"},{"why":"The low-cost car-like robot platform used for the real-world validation experiments.","marker":"[40]"},{"why":"The idealized linear time-varying stochastic differential equation model used for the baseline comparisons.","marker":"[45]"}],"fun_headline_variants":["STELA: one factor graph for nonlinear robot control at 10Hz","Trajectory tracking and steering unified in a single factor graph","Robot adapts controls and estimates path in one graph","Factor graph unifies estimation and control for car-like robots","Nonlinear robots follow plans via one updatable factor graph"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands on the assumption that, on each short segment, acceleration is a function of the control input alone and constant-velocity Euler integration accurately propagates the state; for systems where acceleration also depends on pose or velocity, the dynamics factor used in the graph does not actually represent the system.","fun_headline_variants_meta":{"raw":{"variants":["STELA: one factor graph for nonlinear robot control at 10Hz","Trajectory tracking and steering unified in a single factor graph","Robot adapts controls and estimates path in one graph","Factor graph unifies estimation and control for car-like robots","Nonlinear robots follow plans via one updatable factor graph"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000521,"raw_usage":{"total_tokens":2568,"prompt_tokens":1040,"completion_tokens":1528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1444}},"tokens_in":656,"tokens_out":1528,"duration_ms":11948,"temperature":1.0,"reasoning_tokens":1444,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:37:39.390435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a damped double integrator with known damping coefficient $c$, so the true dynamics are $­\\ddot{q} = -c \\dot{q} + u$, and run STELA with the paper's dynamics factor, which predicts $­\\dot{q}_{i+1} = \\dot{q}_i + f(u_i)\\Delta t$ without the $-c \\dot{q}$ term. If the forward predictions become systematically biased and the success rate falls as $c$ grows, the claim that the representation applies to any dynamical system with first- or second-order state updates is not supported beyond the control-affine form.","supporting_citations":[{"cited_title":"Foundations and Trends® in Robotics 6(1- 2), 1–139 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the factor graph formalism and MAP inference background on which STELA's representation builds."},{"cited_title":"The International Journal of Robotics Research 31(2), 216–235 (2012)","cited_arxiv_id":null,"evidence_quote":"Provides the iSAM2 incremental smoothing and mapping algorithm used for efficient factor graph updates."},{"cited_title":"IEEE Transactions on Robotics 24(6), 1365–1378 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides the earlier iSAM incremental smoothing algorithm that motivates STELA's high-frequency updates."},{"cited_title":"In: 2022 International Conference on Robotics and Automation (ICRA)","cited_arxiv_id":null,"evidence_quote":"The closest prior method, SCATE, which STELA extends and compares against; limited to linear time-varying dynamics and constant duration steps."},{"cited_title":"In: 2020 IEEE International Conference on Robotics and Automation (ICRA)","cited_arxiv_id":null,"evidence_quote":"The asymptotically optimal kinodynamic sampling-based planner whose output initializes the STELA factor graph."},{"cited_title":"Au- tonomous Robots 43, 415–434 (2019)","cited_arxiv_id":null,"evidence_quote":"Prior simultaneous trajectory estimation and planning framework that STELA generalizes beyond Gaussian-process dynamics."},{"cited_title":"The International Journal of Robotics Research 37(11), 1319–1340 (2018)","cited_arxiv_id":null,"evidence_quote":"Source of the obstacle factor definition and of the continuous-time Gaussian process approach that previously limited such methods to holonomic linear systems."},{"cited_title":"IEEE Transactions on Robotics (2024)","cited_arxiv_id":null,"evidence_quote":"The idealized linear time-varying stochastic differential equation model used for the baseline comparisons."}],"review_version":1}