{"id":"817ae9f2-b795-4bd1-b495-2cca125172de","arxiv_id":"2505.05507","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Using a variational integrator inside MPPI rollouts lets the controller plan 4-20 times further ahead, improving balance uptime on underactuated double pendulums.","lead":"A control method for double pendulums swaps the standard numerical integrator inside an MPPI planner for a physics-preserving \"variational\" integrator, allowing longer planning horizons at the same compute cost. In AI Olympics tests it balanced the pendubot and acrobot far longer than the published baselines it was compared with.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) uses the mixed partial D1D2Ld as the Jacobian, but the momentum residual's derivative is D2D2Ld; the one-step correction is not a Newton step, so the variational integrator's accuracy at Δt=0.02 is unsubstantiated.","rationale":"The reader's weakest assumption was that a single Newton correction in the variational solve is sufficient, with no error analysis provided. My concern goes further: the correction as written in Eq. (11) is not a Newton correction at all, because the chosen matrix is the mixed partial D₁D₂L_d rather than the derivative of the residual, D₂D₂L_d. This is an internal inconsistency, not a matter of consensus. If confirmed, it would invalidate the theoretical justification for using a 0.02 s timestep and for calling the method a variational integrator. However, the empirical results are still plausible; the authors might have made a typo in the paper or use an unconventional approximation that happens to work. The appropriate response is therefore to require a correction and an error analysis, which matches the reader's CONDITIONAL verdict. The concrete test would settle whether this concern actually lands. The missing computational-cost measurement is important but less fundamental, since even without that measurement the core algorithmic claim of variational integration is at stake.","tokens_in":6038,"tokens_out":7064,"duration_ms":65566,"concrete_test":"For the pendubot/acrobot dynamics, compute the residual r(q_{k+1}) after one iteration with the paper's A=D₁D₂ versus the exact Newton A=D₂D₂ at Δt=0.02 over random states. Also measure energy drift over 10^4 steps. If the paper's A does not reduce ||r|| and yields energy drift close to explicit Euler, while the exact Newton does, the error is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is internal: the update (12)-(13) is not a Newton step for the equation it claims to solve. The discrete momentum residual is r(q_{k+1}) = p_k - p̂_k = p_k + ∇_{q_{k+1}} L_d(q_k,q_{k+1}) (using p̂_k = -∇_{q_{n+1}}L_d, Eq. (10)). The derivative of r with respect to q_{k+1} is the pure second derivative ∇²_{q_{k+1}} L_d (D₂D₂). Equation (11) instead sets A = ∇_{q_{k+1}}(∇_{q_k}L_d) = D₁D₂, the mixed partial. For the midpoint discrete Lagrangian (7), D₂D₂L_d = (1/Δt)M + (Δt/4)∇²V and D₁D₂L_d = -(1/Δt)M + (Δt/4)∇²V, which differ by 2M/Δt. At Δt=0.02 s this is large. Therefore one application of (12) does not in general drive the momentum error to zero, and the claim that one correction leaves 'negligible' error is not supported by the equations. The symplectic/accuracy argument for using Δt=0.02 rests entirely on this solve; if the solve is wrong, the 4-20x horizon advantage is not attributable to variational integration. This is a concrete correctness risk.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents VIMPPI, a Model Predictive Path Integral (MPPI) controller for underactuated double pendulum systems (pendubot and acrobot) developed for the AI Olympics competition. The core idea is to replace the standard explicit Euler integrator in MPPI rollouts with a variational integrator based on a discrete Euler-Lagrange momentum formulation, allowing much larger rollout timesteps (0.02 s versus 0.001-0.005 s) and hence longer planning horizons 'without additional computational cost.' Additional practical components include linear control interpolation, a disturbance detection and warm-start mechanism, and a JAX implementation. Experiments on an NVIDIA RTX 4070 Ti Super show uptime improvements over several baselines from previous competitions and over MPPI variants using other integrators, with the largest gains on the acrobot.","tokens_in":6385,"tokens_out":4803,"duration_ms":45766,"significance":"If the central claims hold, the paper offers a simple, model-based way to extend the effective planning horizon of sampling-based MPC for underactuated systems, with potential applicability to resource-constrained platforms. The empirical comparison against multiple external baselines (evolsac, AR-EAPO, mcpilco, TVLQR) and alternative integrators is a strength, and the competition setting provides a concrete, reproducible benchmark. The use of JAX for parallel rollouts is also appropriate. However, the theoretical justification of the efficient variational-integrator solve is incomplete: the Jacobian used in the one-step correction appears incorrect, and the 'without additional computational cost' claim is not backed by any timing measurements. These issues directly affect the central contribution and must be addressed before the paper's conclusions can be accepted.","major_comments":[{"comment":"The one-step configuration correction is not a Newton step for the momentum residual it claims to solve. The momentum residual is r(q_{k+1}) = p_k - \\hat{p}_k = p_k + \\nabla_{q_{k+1}} L_d(q_k, q_{k+1}), so its derivative with respect to q_{k+1} is the pure second derivative D_2D_2 L_d. Equation (11) instead sets A = \\nabla_{q_{k+1}}(\\nabla_{q_k} L_d) = D_1D_2 L_d, the mixed partial. For the midpoint discrete Lagrangian (7), D_2D_2L_d = (1/\\Delta t)M + (\\Delta t/4)\\nabla^2 V and D_1D_2L_d = -(1/\\Delta t)M + (\\Delta t/4)\\nabla^2 V, which differ by 2M/\\Delta t. At \\Delta t = 0.02 s this difference is large, so the update (12)-(13) does not in general drive the momentum error to zero. The claim in Section II-B that 'a single error correction iteration is typically sufficient' is therefore unsupported by the equations. Please either use the correct Jacobian D_2D_2L_d or provide a numerical error analysis demonstrating that the iteration converges with the current A at the operating timestep.","section":"Section II-B, Eqs. (10)-(13)"},{"comment":"The central claim that the 4-20x increase in effective planning horizon is achieved 'without additional computational cost' is not substantiated by any computational cost measurement. The paper reports an operating frequency of 500-700 Hz for the full controller, but it does not report wall-clock time, FLOPs, or the number of integration steps for the variational integrator versus explicit Euler at a matched planning horizon and sample count. The variational integrator uses an extra correction step per rollout step, which has a nonzero cost; whether the larger timestep fully offsets this cost is exactly what needs to be measured. Please provide a timing or complexity comparison for VIMPPI against an Euler-based MPPI with equivalent planning horizon and sample size.","section":"Sections III-A and V"},{"comment":"The comparison against other integrators under 'the same setup horizon conditions' is not well defined. If all integrators are run with the same rollout timestep of 0.02 s, explicit Euler is likely unstable or highly inaccurate, making the comparison trivially favorable to the variational integrator. If each integrator instead uses a different stable timestep, then the 4-20x horizon increase is conflated with the choice of timestep, and the computational cost of each integrator for the same planning horizon is not controlled. Please specify the exact conditions (timestep, horizon, number of samples) used for each integrator and report the resulting planning accuracy and wall-clock time.","section":"Section IV"}],"minor_comments":[{"comment":"The stage cost uses \\Sigma_u^{-1}, while Table I lists \\sigma = 0.2 I_2; please clarify the relation between \\sigma and \\Sigma_u and specify the units of the noise variance.","section":"Equation (4) and Table I"},{"comment":"The phrase 'first application of such an approach' in the Introduction is already qualified with 'To our knowledge,' but the abstract states it more categorically; please align the wording in both places.","section":"Abstract and Introduction"},{"comment":"The tables report 'Swingups' with mean and standard deviation, but the text emphasizes uptime; please clarify whether swingups count successful transitions or include failed attempts, and consider adding a statistical significance test between the top two controllers.","section":"Tables II and III"},{"comment":"The notation for the momentum residual uses both q_n and q_{n+1} in Eq. (10) and q_k and q_{k+1} in Eqs. (11)-(13); please unify the index notation to avoid confusion.","section":"Section II-B, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competition-oriented manuscript whose central contribution rests on the variational integrator's accuracy and cheapness at large timesteps. The wrong Jacobian in the one-step correction is a concrete, fixable flaw, but it undermines the current justification of the method. The missing computational cost measurement also weakens the headline claim. If the authors can correct the integrator step and add the missing measurements, the paper would be significantly stronger; however, in its current form the theoretical support is not sufficient to accept the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of the VIMPPI paper. The core idea is straightforward and appealing: replace the usual explicit Euler rollout in MPPI with a variational integrator, so you can take 20 ms rollout steps instead of 1–5 ms and get a longer planning horizon for the same number of steps. The combination appears to be new, and the authors credit West and Williams appropriately. The results on the AI Olympics pendubot/acrobot benchmark are strong: VIMPPI beats the next-best controller by 8–11 s uptime with lower variance, and runs at 500–700 Hz. That is a solid external benchmark result.\n\nThe soft spots are real, though. The headline claim \"without additional computational cost\" is not backed by any cost measurement. The correction step in Eqs. (11)–(12) requires forming and inverting a matrix, which is not obviously free, and the paper never reports per-rollout or wall-clock time. The integrator comparison also uses \"the same setup horizon conditions,\" meaning the explicit and implicit baselines are run at the same large timestep as the variational integrator. That stacks the deck against them; a fair comparison would run each integrator at its stable timestep and compare cost for equal horizon.\n\nA more technical worry: the Jacobian in Eq. (11) is defined as A = ∇_{q_{k+1}}(∇_{q_k} L_d), a mixed partial. If p̂_k in Eq. (10) is the momentum estimate from the pair (q_k, q_{k+1}), then the residual in Eq. (12) is p_k + D_2 L_d(q_k, q_{k+1}), whose derivative w.r.t. q_{k+1} is D_2D_2 L_d, not the mixed partial. For the midpoint discrete Lagrangian these differ by 2M/Δt. So the update (12)–(13) is not the Newton step the text describes, and the claim that one iteration leaves negligible error is not supported by the equations. The authors need to either fix the Jacobian or provide a direct accuracy analysis of the integrator at Δt = 0.02. Without that, the 4–20x horizon advantage is not clearly attributable to variational integration.\n\nThe paper is still worth refereeing: the empirical gains are substantial, the benchmark is external, and the idea is timely. But those two gaps—cost measurement and the integrator's derivation—need to be resolved. I'd send it out with a request for major revision. Sharing the JAX code would also help; the paper doesn't mention code availability, which is a small but real reproducibility gap.","headline":"Useful combination of variational integration and MPPI with strong benchmark results, but the no-cost claim is unmeasured and the correction-step Jacobian in Eq. (11) looks inconsistent with the residual in Eq. (12).","tokens_in":6914,"tokens_out":8631,"would_cite":false,"duration_ms":73850,"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":"Swapping in a variational integrator extends MPPI planning horizons 4–20x and wins longer balance times on underactuated pendulums.","keywords":["model predictive path integral control","variational integrator","symplectic integration","underactuated double pendulum","pendubot","acrobot","sampling-based control","real-time control"],"falsifier":"Run VIMPPI's 0.02 s rollout against a high-accuracy reference (e.g., RK4 or a variational integrator converged to tight tolerance at $\\Delta t = 10^{-4}$ s) over the full 20-step horizon and measure state error and energy drift. If the single-correction rollout drifts faster than an explicit Euler rollout at 0.001 s given the same wall-clock compute budget, or if reducing the sample count to offset the larger timestep loses the reported uptime margins, the central 'longer horizon at no extra cost' claim would be refuted.","tokens_in":5841,"feed_emoji":"⚙️","tokens_out":5518,"duration_ms":47438,"temperature":0.7,"pith_summary":"This paper argues that the bottleneck in Model Predictive Path Integral (MPPI) control is not the optimizer but the numerical integrator used in its rollouts. Replacing the standard explicit Euler integrator with a variational (symplectic) integrator lets each rollout step jump from 0.001–0.005 s to 0.02 s or more while staying accurate, stretching the effective planning horizon by 4–20x at the same compute budget. On two underactuated double pendulums (pendubot and acrobot) the resulting controller, VIMPPI, holds the upper equilibrium for up to 48 s, outperforming prior learning-based baselines and MPPI variants with other integrators. The authors position this as a general lesson: in sampling-based control of well-modeled mechanical systems, the choice of integrator deserves as much attention as the sampling scheme.","feed_headline":"Variational integrator gives MPPI 4–20x longer planning horizons","feed_subtitle":"VIMPPI holds pendubot and acrobot upright up to 48 s, beating best baselines by 8-17 s.","key_machinery":"The load-bearing object is the discrete Euler–Lagrange residual in momentum form, paired with a midpoint discrete Lagrangian $L_d(q_n,q_{n+1}) = L((q_n+q_{n+1})/2, (q_{n+1}-q_n)/\\Delta t)\\Delta t$. Instead of solving the implicit update by rootfinding, the integrator takes an explicit Euler guess and applies one correction $\\Delta \\hat q_i = A^{-1}(p_k - \\hat p_k)$, where $p_k$ is the momentum from the known velocity, $\\hat p_k$ is the momentum read off the discrete Lagrangian, and $A$ is the Jacobian of the discrete Lagrangian gradient. This single correction is what keeps the rollout accurate and effectively symplectic at $\\Delta t = 0.02$ s, and it is what converts a larger timestep into a longer planning horizon for the same number of rollouts.","core_discovery":"VIMPPI's central claim is that a variational integrator—one that discretizes Hamilton's principle rather than the equations of motion—can serve as the rollout engine of MPPI, making the planning horizon 4–20 times longer without extra computational cost. Concretely, the paper replaces explicit Euler integration with a discrete Euler–Lagrange solve in momentum form: after an Euler guess for the next configuration, a single Newton-type correction based on the mismatch between true momentum and the momentum implied by the discrete Lagrangian brings the rollout onto the symplectic flow. Because energy and momentum drift are controlled at large timesteps, the cost-to-go computed over the horizon reflects true dynamics rather than numerical artifact. The paper reports that with a 0.02 s rollout timestep and a 20-step horizon at 500–700 Hz, VIMPPI maintains the unstable equilibrium for $48.03 \\pm 0.9$ s on the pendubot and $38.6 \\pm 2.51$ s on the acrobot, beating the next-best controllers by roughly 8 and 11 s respectively.","pith_inferences":["A testable extension is to run the same VIMPPI update on a damped or actuated system with non-conservative terms; variational integrators do not preserve energy there, so the claimed 4–20x advantage would likely shrink unless dissipation is folded into the discrete Lagrangian.","The one-step Newton correction is presented as empirically sufficient; measuring the contraction rate of the correction map as a function of $\\Delta t$ would tell whether the method's accuracy guarantee extends beyond the two benchmark systems.","Since MPPI's importance weights depend on rollout costs, a more accurate rollout integrator may also reduce estimator variance for a fixed sample count, which could let practitioners cut the 4096 samples and recover some of the computational budget.","If the horizon extension is as large as claimed, VIMPPI-type rollouts could serve as a cheap forward model for model-based reinforcement learning, replacing learned dynamics for systems whose physics are known."],"forward_implications":["If the integrator is the bottleneck, then any sampling-based predictive controller for a well-modeled mechanical system can gain a 4–20x horizon extension by swapping its rollout integrator, with no change to the sampling or cost structure.","At equal horizon, VIMPPI's larger timestep can lower the number of rollout steps per plan, reducing per-iteration compute and allowing higher control rates (500–700 Hz here) on the same hardware.","The reported uptime margins (about 8 s on the pendubot, 11 s on the acrobot over the next-best controller) follow from the longer lookahead, so the benefit should grow for systems with limited control authority, where prediction quality matters more.","Because the variational integrator preserves conservative structure, the approach should transfer to other underactuated platforms whose dynamics are Lagrangian, such as multi-link manipulators, quadrupeds, and humanoids."],"supporting_citations":[{"why":"Supplies the MPPI update law and stochastic optimal control formulation that the paper builds on.","marker":"[10]"},{"why":"Foundation for the discrete Euler–Lagrange momentum-form variational integrator and its correction scheme.","marker":"[11]"},{"why":"Initial MPPI application to aggressive vehicle control that motivates sampling-based planning.","marker":"[9]"},{"why":"The parallel-computation library that makes 4096 rollouts at 500–700 Hz feasible in real time.","marker":"[12]"},{"why":"Defines the pendubot/acrobot benchmark platform and upper-equilibrium task used in evaluation.","marker":"[2]"},{"why":"The baseline controller used both as comparison and as warm-start initialization for disturbance recovery.","marker":"[5]"}],"fun_headline_variants":["Symplectic integrator extends MPPI horizons 20x, holds pendubot 48s","VIMPPI: variational integration gives MPPI 4–20x longer planning at no extra cost","Variational MPPI: up to 20x longer horizons, 48s pendubot hold","VIMPPI: symplectic rollouts extend planning horizon 20x, hold acrobot 38s"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that one Newton-style correction in the discrete Euler–Lagrange solve keeps the 0.02 s rollout accurate enough that the cost-to-go faithfully represents the true dynamics over the planning horizon; the paper reports this by observation but gives no error analysis, and if the single correction is inaccurate at large timesteps the horizon advantage over smaller-timestep integrators disappears.","fun_headline_variants_meta":{"raw":{"variants":["Symplectic integrator extends MPPI horizons 20x, holds pendubot 48s","VIMPPI: variational integration gives MPPI 4–20x longer planning at no extra cost","Variational MPPI: up to 20x longer horizons, 48s pendubot hold","VIMPPI: symplectic rollouts extend planning horizon 20x, hold acrobot 38s"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001066,"raw_usage":{"total_tokens":4416,"prompt_tokens":843,"completion_tokens":3573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":3463}},"tokens_in":459,"tokens_out":3573,"duration_ms":25301,"temperature":1.0,"reasoning_tokens":3463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:34:03.744485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VIMPPI's 0.02 s rollout against a high-accuracy reference (e.g., RK4 or a variational integrator converged to tight tolerance at $\\Delta t = 10^{-4}$ s) over the full 20-step horizon and measure state error and energy drift. If the single-correction rollout drifts faster than an explicit Euler rollout at 0.001 s given the same wall-clock compute budget, or if reducing the sample count to offset the larger timestep loses the reported uptime margins, the central 'longer horizon at no extra cost' claim would be refuted.","supporting_citations":[{"cited_title":"Variational integrators,","cited_arxiv_id":null,"evidence_quote":"Foundation for the discrete Euler–Lagrange momentum-form variational integrator and its correction scheme."},{"cited_title":"Open source dual-purpose acrobot and pendubot platform: Bench- marking control algorithms for underactuated robotics,","cited_arxiv_id":null,"evidence_quote":"Defines the pendubot/acrobot benchmark platform and upper-equilibrium task used in evaluation."},{"cited_title":"Average-Reward Maximum Entropy Reinforcement Learning for Underactuated Double Pendulum Tasks","cited_arxiv_id":"2409.08938","evidence_quote":"The baseline controller used both as comparison and as warm-start initialization for disturbance recovery."}],"review_version":1}