{"id":"12fb2ea2-2e9e-4901-abfc-6f36fdbb677e","arxiv_id":"2411.18295","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A closed-form least-squares optimization selects a knee torsion spring's stiffness and equilibrium position, cutting simulated motor energy to under 4% of baseline in the same cyclic task used for fitting.","lead":"The authors derive a formula for the best stiffness and rest angle of a torsion spring added to a legged robot knee, and test it in a Gazebo simulation. They report large cuts in a motor energy metric for cyclic motion, but the spring is fitted to the very same motion used for evaluation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central energy-reduction claim rests on a torque-squared proxy (Eq. 1) that omits mechanical work and spring-adjustment energy, so Table I's Ea/E0 ratios may not represent total energy consumption.","rationale":"The reader's weakest assumption concerns in-sample validation: parameters are fitted to the baseline trajectory and then evaluated on the same trajectory with the spring. That concern is real but not the most load-bearing, because for a fixed cyclic task the same-trajectory fit is actually the relevant objective: if the trajectory is preserved, the least-squares solution is the global minimizer of the stated squared-torque metric. The more serious issue is that the stated metric itself is not a complete measure of energy consumption. Eq. 1 ignores the mechanical work term of the motor and the energy consumed by the spring-adjustment mechanism, so Table I does not support the paper's broad conclusion about overall energy reduction. This concern is testable with a concrete motor model and would, if confirmed, require the authors to either revise the claim or augment the experiments. I therefore keep the verdict conditional rather than rejecting outright, since the algebraic derivation and simulation pipeline can be repaired by reporting a more faithful energy measure.","tokens_in":6305,"tokens_out":10529,"duration_ms":107764,"concrete_test":"Re-run the experiment for at least one row of Table I (e.g., m=4.1, T=1.88, A=0.05, h0=0.2) using a full motor electrical model: log joint torque and velocity, set i = tau / K_t, and compute E_total = integral(R i^2 + k_e i omega) dt for both baseline and spring-equipped runs, under two assumptions: ideal regeneration and non-regenerative braking. Also add the measured energy drawn by the equilibrium-adjustment servo. If E_total_spring / E_total_baseline is substantially higher than the Table I ratio, the headline overstates energy savings.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline result, Ea/E0 ratios as low as 0.15%, is obtained by minimizing Eq. 2, which is a sum of squared actuator torques. Eq. 1 asserts that total actuator energy is E = K * sum(tau_i^2 * dt), but this is a resistive-loss-only proxy. For a real DC motor, electrical power is approximately i^2 R + i k_e omega, where current i is proportional to torque. The mechanical term i k_e omega does not vanish during a cyclic task; it represents positive work drawn from the battery and, in non-regenerative drives, negative work that is dissipated rather than returned. The model also omits the energy consumed by the Dynamixel servo and worm gear that adjust and maintain the spring's equilibrium position, which is part of the proposed adaptive mechanism. Therefore, a spring that cancels nearly all torque-squared could still leave a large fraction of total battery energy unchanged or even increase it, depending on the motor constants and drive electronics. The paper's claim that 'overall energy consumption could be considerably reduced' is thus not established by Table I; at best, the table shows reduction of the squared-torque loss proxy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive torsion spring mechanism for legged robots, with an adjustable equilibrium position and stiffness, intended to reduce actuator energy consumption during cyclic tasks. The main contribution is a closed-form derivation of optimal spring parameters (stiffness μ* and equilibrium position α0*) that minimize a torque-squared energy proxy (Eqs. 1–4). The method is evaluated in a Gazebo simulation of a 3-DoF leg controlled by a PD controller, comparing a baseline run with a run where a torsion spring with the derived parameters is added. Table I reports energy ratios Ea/E0 between 0.15% and 3.9% over variations in mass, frequency, amplitude, and starting height, and the paper concludes that the overall energy consumption can be considerably reduced.","tokens_in":6498,"tokens_out":4210,"duration_ms":37372,"significance":"If the reported quantitative reductions were valid, the closed-form formulas would provide a straightforward and computationally cheap method for sizing parallel torsion springs in cyclic robotic tasks. The derivation of Eq. (4) is mathematically correct, and the authors make their code and simulation setup available, which supports reproducibility. However, the central claim of significant energy reduction is not established by the evidence presented: the energy metric used is a resistive-loss proxy, the evaluation is in-sample, and the energy cost of the adaptive mechanism itself is omitted. These issues are load-bearing for the paper's main conclusion, so the significance of the contribution as currently stated is limited.","major_comments":[{"comment":"The energy metric E = K Σ τ_i² Δt is a torque-squared proxy (proportional to resistive losses in a DC motor) and omits the mechanical power term i k_e ω. In a non-regenerative drive, negative mechanical work is dissipated rather than recovered, so the true electrical energy is not proportional to the sum of squared torques. Consequently, the E_a/E0 ratios reported in Table I do not represent total or even electrical energy consumption, and the conclusion in Section V that 'the overall energy consumption could be considerably reduced' is not supported by this metric.","section":"Section III-B, Eq. (1) and Table I"},{"comment":"The optimal spring parameters are computed from the baseline trajectory without the spring, and then energy is measured on that same trajectory with the spring. Since Eq. (4) is the least-squares solution for fitting μ(α - α0) to the baseline torque τ, the ratio E_a/E0 equals 1 - R² of that regression. The reported values (0.15% to 3.9%) therefore quantify the in-sample goodness of fit of a linear spring model to the baseline torque-angle data, not an energy saving. A held-out trajectory, a different gait, or cross-validation is necessary to support any predictive claim about energy reduction.","section":"Section IV-A and Table I"},{"comment":"The paper assumes that the actuator torque as a function of joint angle is unchanged by adding the spring, so that parameters fitted to the baseline remain optimal for the spring-equipped system. However, the spring alters the plant dynamics (it adds torque that changes the acceleration and the closed-loop tracking error), so the actuator torque during the spring run is not necessarily τ_i - μ(α_i - α0). The paper does not verify whether the same parameters would be obtained if the optimization were repeated with the spring present, nor does it quantify trajectory deviations between the baseline and spring-equipped runs.","section":"Section IV-A and Section IV-B"},{"comment":"The energy consumed by the Dynamixel servo and worm gear that adjust and maintain the spring's equilibrium position is not included in Eq. (2) or in Table I. Since this actuation is an integral part of the proposed adaptive mechanism, a fair comparison of total system energy consumption must account for it; otherwise the reported savings could be substantially offset or even reversed.","section":"Section VI and Figs. 3–4"}],"minor_comments":[{"comment":"The keyword 'Adaptive Comlpliance' appears to be a typo; it should be 'Adaptive Compliance'.","section":"Keywords"},{"comment":"The phrase 'Part if the recent research' in Section I should be 'Part of the recent research'.","section":"Introduction"},{"comment":"The sentence 'The proposed system system relies on a torsion spring' contains a duplicated word 'system'.","section":"Section I"},{"comment":"The conclusion states 'in the experiments with a physical stand the reduction in power consumption will be not as fascinating', but no physical experiments are reported in this paper; the text should clarify that these are simulation results and that a physical test is planned for future work.","section":"Section V"},{"comment":"The conditions described in the Figure 2 caption (e.g., 'Frequency increased from 100Hz up to 200Hz') do not match the corresponding entries in Table I (which lists periods 0.94, 1.88, 3.77 s). Please align the figure descriptions with the experimental parameters.","section":"Figure 2 and Table I"},{"comment":"The table is missing units and clearly defined column headers: μ* is presumably in N·m/rad, α0* in rad, E0 and Ea in (N·m)²·s (or a related unit), and 'h0' is the initial height in meters. Adding units would improve interpretability.","section":"Table I"},{"comment":"Some references are incomplete or inconsistently formatted, e.g., [3] is a URL-only entry and [17] omits the full author list. Please check journal/conference standards.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the in-sample evaluation: the reported energy reductions are essentially the R² of a least-squares fit, not measured energy savings. If the authors re-run experiments with held-out tasks and a more complete energy model, the paper could become publishable. There is also some overlap with existing adjustable-equilibrium parallel elastic actuators (e.g., [10]); the novelty claim should be more carefully positioned. The closed-form derivation itself is correct and may be useful to the community, so I would recommend a major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the math is right, the mechanism is plausible, but Table I does not establish overall energy consumption reduction. The paper fits spring parameters to a baseline trajectory, then measures the torque-squared metric on that same trajectory. The 0.15%–3.9% ratios are essentially the residual variance of a linear fit; they tell you how well a linear spring approximates the baseline torque curve, not how much battery energy you save.\n\nWhat is genuinely useful: the closed-form solution (Eq. 4) is the correct least-squares optimum for the stated proxy, and the torsion-spring-with-worm-gear hardware concept is simple and worth testing. The simulation covers several masses, frequencies, amplitudes, and starting heights, and the code is on GitHub. That is real, reproducible work for what it is.\n\nThe soft spots are not hiding. First, Eq. 1 is a resistive-loss proxy; it omits the mechanical power term i*k_e*omega, which does not vanish in cyclic motion, and it omits the energy consumed by the Dynamixel servo and worm gear that adjust the spring. A spring that cancels most of the torque-squared could still leave most of the battery consumption intact. Second, the evaluation is entirely in-sample: the same trajectory generates the fit and the metric. No held-out trajectories, no task variation, no repeat runs, no error bars. The word adaptive overstates it—experiments are offline parameter optimization, not closed-loop adaptation. Third, the claim in Section V that overall energy consumption could be considerably reduced goes beyond what the metric shows.\n\nThat said, none of this is fatal to the core idea. The derivation is correct, the hardware concept is worth pursuing, and the deficiencies are addressable with a proper energy model (including mechanical work and actuation cost), held-out evaluation, and repeated trials.\n\nI would send this to a serious referee, not desk reject it. The reader is right that it is incremental relative to refs 6 and 10, but the torsion spring mechanism and the simplicity of the closed-form are enough to justify a review, with major revision expected. For my own work, I would not cite the energy numbers, but I might cite the mechanism concept if I built on it.","headline":"A clean closed-form least-squares fit for parallel spring parameters, but the headline energy savings are in-sample residuals of a torque-squared proxy, not measured total energy.","tokens_in":7083,"tokens_out":2046,"would_cite":false,"duration_ms":18330,"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":"A torsion spring with closed-form optimal stiffness and equilibrium position can reduce simulated energy consumption of a cyclic legged-robot task to as little as 0.15% of baseline.","keywords":["adaptive compliance","torsion spring","energy efficiency","legged robots","parallel elastic actuator","closed-form optimization","Gazebo simulation","cyclic motion"],"falsifier":"Run the same cyclic motion in the simulator (or on the physical stand) with the spring installed using the baseline-optimal $\\mu^*$ and $\\alpha_0^*$, then record the actual actuator torque after the spring is in place; if the measured energy ratio differs substantially from the predicted $E_a/E_0$, the one-step closed-form prediction is invalid. Alternatively, allow the controller to re-optimize the trajectory with the spring present: if this yields markedly lower energy than the paper's predicted ratio, the two-step procedure is suboptimal.","tokens_in":6084,"feed_emoji":"🤖","tokens_out":6131,"duration_ms":49962,"temperature":0.7,"pith_summary":"The paper sets out to show that a single torsion spring mounted in parallel with the knee actuator of a legged robot can be tuned, in one closed-form step, to compensate most of the torque the motor would otherwise produce during a cyclic motion. The tuning rule minimizes the sum of squared actuator torques over a recorded trajectory, yielding explicit formulas for the spring's stiffness and equilibrium position. In Gazebo simulations of a three-degree-of-freedom leg performing vertical oscillations, the spring-equipped system consumes as little as 0.15% of the baseline energy, with the reduction holding across changes in mass, frequency, amplitude, and starting height. The point of the work is that a mechanically simple adjustment could extend battery life and reduce servo wear without iterative learning or online search.","feed_headline":"One spring formula slashes robot-leg energy to 0.15%","feed_subtitle":"Optimal stiffness and equilibrium position from one baseline trajectory cancel nearly all knee-motor torque in simulation.","key_machinery":"The central object is the quadratic energy functional over one motion cycle, $E = K\\sum_{i=1}^T (\\tau_i - \\mu(\\alpha_i - \\alpha_0))^2 \\Delta t$, where $\\tau_i$ is the actuator torque, $\\alpha_i$ the joint angle, $\\mu$ the spring stiffness, and $\\alpha_0$ the spring equilibrium position. Treating the spring torque as a linear regressor for the actuator torque turns parameter optimization into a linear least-squares problem, and Equation 4 gives the closed-form normal-equation solution for $\\mu^*$ and $\\alpha_0^*$. This is what carries the argument: no iterative search, no learning, just a direct computation from a single recorded trajectory.","core_discovery":"For a repeated cyclic task, the energy-optimal linear torsion spring is the one whose torque, $\\mu(\\alpha - \\alpha_0)$, is the least-squares fit to the actuator torque profile over the cycle. The paper derives the closed-form solution for $\\mu$ and $\\alpha_0$ by differentiating the energy functional $E = K\\sum_{i=1}^T (\\tau_i - \\mu(\\alpha_i - \\alpha_0))^2 \\Delta t$ and solving the resulting normal equations. Applying these parameters in simulation cancels nearly all of the knee motor's torque, yielding energy ratios $E_a/E_0$ from 0.15% to 3.9% across the tested conditions. The authors present this as evidence that an adaptive torsion spring with adjustable equilibrium position can considerably reduce energy consumption in cyclic legged locomotion.","pith_inferences":["The closed-form solution is mathematically a least-squares projection of the torque profile onto the affine line spanned by the joint angle, so its 'optimality' is only with respect to the squared-torque cost and the trajectory recorded without the spring; the paper does not prove that the spring leaves the torque-angle relationship unchanged.","A natural test the authors do not run: replan the leg trajectory with the spring in place and re-optimize. Their one-shot baseline-to-spring procedure may be conservative; a joint optimization could yield even lower energy or expose that the spring's added inertia shifts the optimal parameters.","The same derivation applies to any cyclic joint torque, not just the knee in this stand setup, so the formula could transfer to other joints or gaits; validating it on a walking gait with multiple coupled joints would be the next step.","In a physical system, friction and the servo that adjusts the preload consume energy themselves; the reported ratios are for the ideal spring model, and a hardware test would need to account for these overheads before claiming net savings."],"forward_implications":["In simulation, the spring reduces energy consumption to between 0.15% and 3.9% of baseline across varying mass, frequency, amplitude, and starting height, implying near-total torque cancellation in the tested cyclic task.","The optimal parameters depend on the task conditions (stiffness ranges from 6.07 to 17.1, equilibrium position from -1.4 to -2.84), so the adjustable equilibrium position is essential rather than decorative.","Lower actuator torque implies less heat generation and less mechanical wear on the servos, which the paper cites as a direct practical benefit.","Because the parameters come from a closed-form solution, the approach is computationally light and could run in real time or be set once per task without training data.","The proposed mechanism, a torsion spring with a worm-gear-driven preload adjustment, extends the same idea to hardware, though the current evidence is entirely from the Gazebo simulator."],"supporting_citations":[{"why":"Gazebo joint torsional spring plugin used to simulate the spring's torque in all experiments.","marker":"[15]"},{"why":"Prior work on adapting equilibrium position of parallel compliance in cyclic tasks, which this paper extends by giving a closed-form solution.","marker":"[6]"},{"why":"Learning-based design of parallel-elastic springs for quadrupeds, the iterative approach the paper contrasts with its direct calculation.","marker":"[1]"},{"why":"Adjustable-equilibrium parallel elastic actuator design that motivates the variable preload mechanism.","marker":"[10]"},{"why":"Physical test bench and electronics for the leg, the intended hardware platform for future validation.","marker":"[17]"}],"fun_headline_variants":["Tuning one spring cuts robot knee energy to 0.15%","Optimal spring stiffness and offset slash leg energy to 0.15%","Adaptive spring formula slashes legged robot energy to 0.15%","Closed-form spring fit cuts legged robot energy to 0.15%","Spring equilibrium and stiffness formula cuts robot energy to 0.15%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The spring parameters are computed from a trajectory recorded without the spring, and the paper assumes the actuator torque at each joint angle stays the same once the spring is added, even though the spring changes the load dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Tuning one spring cuts robot knee energy to 0.15%","Optimal spring stiffness and offset slash leg energy to 0.15%","Adaptive spring formula slashes legged robot energy to 0.15%","Closed-form spring fit cuts legged robot energy to 0.15%","Spring equilibrium and stiffness formula cuts robot energy to 0.15%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2592,"prompt_tokens":792,"completion_tokens":1800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":1714}},"tokens_in":408,"tokens_out":1800,"duration_ms":13385,"temperature":1.0,"reasoning_tokens":1714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:19:15.871987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same cyclic motion in the simulator (or on the physical stand) with the spring installed using the baseline-optimal $\\mu^*$ and $\\alpha_0^*$, then record the actual actuator torque after the spring is in place; if the measured energy ratio differs substantially from the predicted $E_a/E_0$, the one-step closed-form prediction is invalid. Alternatively, allow the controller to re-optimize the trajectory with the spring present: if this yields markedly lower energy than the paper's predicted ratio, the two-step procedure is suboptimal.","supporting_citations":[{"cited_title":"Gazebo joint torsional spring plugin,","cited_arxiv_id":null,"evidence_quote":"Gazebo joint torsional spring plugin used to simulate the spring's torque in all experiments."},{"cited_title":"Adaptation in variable parallel compliance: Towards energy efficiency in cyclic tasks,","cited_arxiv_id":null,"evidence_quote":"Prior work on adapting equilibrium position of parallel compliance in cyclic tasks, which this paper extends by giving a closed-form solution."},{"cited_title":"Learning-based design and control for quadrupedal robots with parallel-elastic actuators,","cited_arxiv_id":null,"evidence_quote":"Learning-based design of parallel-elastic springs for quadrupeds, the iterative approach the paper contrasts with its direct calculation."},{"cited_title":"Design of a Parallel Elastic Actuator with a Continuously-Adjustable Equilibrium Position","cited_arxiv_id":"2212.07601","evidence_quote":"Adjustable-equilibrium parallel elastic actuator design that motivates the variable preload mechanism."},{"cited_title":"HyperSurf: Quadruped Robot Leg Capable of Surface Recognition with GRU and Real-to-Sim Transferring","cited_arxiv_id":"2407.15622","evidence_quote":"Physical test bench and electronics for the leg, the intended hardware platform for future validation."}],"review_version":1}