Pith. sign in

REVIEW 3 major objections 5 minor 24 references

A single offline-trained network hard-encodes optimality conditions and recovers free-final-time energy-optimal lunar descent over a full initial-state envelope without precomputed trajectories.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 07:21 UTC pith:I3IGE7KN

load-bearing objection Solid free-final-time OINN specialization with hard BCs, closed-form PMP control, residual-to-error bounds, and credible open-loop agreement vs bvp4c; the closed-loop deployment claim outruns the open-loop theory and tests. the 3 major comments →

arxiv 2607.02741 v1 pith:I3IGE7KN submitted 2026-07-02 math.OC cs.LGcs.SYeess.SY

Optimality-Informed Neural Networks for Lunar Landing Trajectory Optimization

Zhenbo Wang This is my paper
classification math.OC cs.LGcs.SYeess.SY MSC 49M3749N9068T0793C95
keywords optimality-informed neural networkslunar powered descentPontryagin minimum principlefree final timephysics residual trainingtrajectory optimizationreal-time guidance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Landing a spacecraft on the Moon with little propellant, from an uncertain starting state, still requires either fragile numerical solves or large datasets of optimal examples. This paper claims that neither is necessary if the network architecture itself is forced to obey the necessary conditions of optimal control. Every boundary and free-time transversality condition is satisfied exactly by construction, the closed-form optimal thrust law is substituted rather than learned, and the remaining state and costate outputs are trained only against residuals of those conditions, with no expert trajectories. The resulting policy is frozen once, evaluated at fixed small cost, and is shown to track independent boundary-value solutions across the operating envelope while keeping dynamics residuals small. A reader who cares about real-time, certifiable planetary guidance would care because the method turns the classical necessary conditions into both the training signal and the safety architecture.

Core claim

An Optimality-Informed Neural Network, with hard-encoded boundary and transversality conditions and the closed-form Pontryagin thrust law substituted, can be trained solely on physics residuals of the necessary conditions and then serve, without retraining, as an approximate energy-optimal free-final-time feedback law for every initial state inside a bounded lunar powered-descent envelope, agreeing closely with independently solved indirect-method solutions and maintaining small residuals throughout that envelope.

What carries the argument

The OINN architecture: a condition encoder that outputs affine velocity-costate parameters and free final time, a main trunk with bubble-function hard constraints that force every boundary and transversality condition to hold exactly, closed-form substitution of the Pontryagin-optimal thrust, and a pure residual loss on the remaining state, costate, Bellman, and Hamiltonian conditions.

Load-bearing premise

The explicit touchdown-error bound is proved for the open-loop trajectory generated by the network's own thrust history under the true dynamics, not for a closed-loop feedback law under model or sensor error, and free-time accuracy further needs the Hamiltonian zero at the true final time to be simple and transversal.

What would settle it

At any of the six named envelope points, re-solve the indirect boundary-value problem from a cold start (or with a different collocation mesh) and check whether the frozen OINN still matches flight time within a few percent and final mass within a fraction of a percent; or, under modest mass or gravity mismatch, integrate the open-loop thrust history and test whether the predicted residual-to-touchdown bound of Theorem 5.3 is violated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes an Optimality-Informed Neural Network (OINN) for energy-optimal, free-final-time lunar powered descent from any initial state in a bounded envelope Ω to a fixed soft-landing site. Building on a prior PMP/HJB-informed framework, it hard-encodes all boundary and transversality conditions into the architecture (bubble functions, softplus final-time map, affine velocity costate), substitutes the closed-form Pontryagin thrust magnitude and direction rather than learning them, and trains only state, mass-costate, and an auxiliary value head against a residual loss of the necessary conditions, with no precomputed optimal trajectories. A theoretical section gives an Adam stationarity guarantee, a Grönwall open-loop residual-to-touchdown bound, a free-time sensitivity estimate, and fixed onboard FLOP/memory counts. Numerics compare one frozen network to independent bvp4c solutions at six envelope points and report residual statistics on eighty Monte Carlo draws.

Significance. If the claims hold as stated, the work is a useful specialization of physics/optimality-informed networks to free-final-time fixed-terminal powered descent: hard architectural satisfaction of BCs/transversality (Lemma 5.2), closed-form PMP control substitution, and training without expert trajectory data are concrete strengths relative to pure imitation or RL guidance. The fixed ~10^4-FLOP forward pass and tens-of-KiB memory footprint are genuinely attractive for onboard use. The residual-to-error bound (Theorem 5.3) and free-time sensitivity remark give a clearer link from training residual to touchdown metrics than is common in PINN-style guidance papers. The contribution is incremental relative to the cited under-review framework [18] and to existing convex and neural landing guidance, but the free-time/fixed-terminal specialization and the open-loop residual analysis are of genuine interest to the optimal-control and GNC communities.

major comments (3)
  1. [§2 eq. (13); §5 Thm 5.3, Rem. 5.1–5.2; abstract; §7] Central deployment claim vs. what is proved and tested. Abstract, §2 (eq. 13), §5, and §7 present a single offline-trained network as a real-time-deployable feedback/guidance law over Ω. The architecture (Fig. 1; eqs. 24–35) and Theorem 5.3 only produce and certify an open-loop time history for a fixed initial state: the Grönwall bound (54) and free-time estimate (55) apply to the trajectory obtained by feeding the network’s own thrust history into the true dynamics from that IC. Lemma 5.2 hard-encodes terminal BCs only on the network’s internal state output, not on a trajectory under mid-course state error, model mismatch, or sensor noise. The six-point bvp4c comparisons and eighty-point Monte Carlo (§6) are likewise pure open-loop residual evaluations. Either (i) rephrase the claim consistently as open-loop trajectory generation parameterized by IC, with closed-loop use left to future
  2. [§5 Assumptions 1 and 3; Thm 5.3; Rem. 5.2; §6] Assumptions that load-bear the accuracy claims are stated but not checked. Theorem 5.3 needs Assumption 1 (m(t)≥m_dry everywhere) for the Lipschitz constant L_f; Remark 5.2 needs Assumption 3 (simple transversal zero of H* at the true t_f). Neither is verified numerically (e.g., min mass over MC draws; sign/magnitude of ∂H*/∂t_f at the predicted root). Without that, the translation from training residual to touchdown and flight-time error remains conditional. Please report these diagnostics or weaken the accuracy claims accordingly.
  3. [Abstract; §1; §4 eqs. (31),(40); Rem. 5.3] HJB language vs. actual role of V. The abstract and introduction emphasize jointly embedding PMP and HJB, yet §4 and Remark 5.3 state that V is only an auxiliary Bellman-consistency diagnostic with residual weight w_V=0.01 and no effect on the commanded control. That is fine technically, but the framing overstates the HJB content relative to a pure PMP residual network. Align the abstract/intro claims with the architecture actually trained.
minor comments (5)
  1. [§2 eq. (11); Rem. 2.2] Energy cost J=∫(T²+ρ)dt is less standard for landing than fuel-optimal (min ∫T or max m_f). Remark 2.2 explains free-time well-posedness, but a short comparison to fuel-optimal structure (and why the closed-form clip law is retained) would help readers place the problem.
  2. [§4 Adaptive Scale Constants; Table 2] Free parameters (ρ, loss weights w_V=w_H=0.01 vs dynamics weights=1, all scale constants s_*, network widths, K, η, h, t_min_f) are numerous. A brief sensitivity note or justification beyond dimensional analysis would strengthen reproducibility.
  3. [§6 Comparison with the Indirect Method] Figures 2–7 are described in text but the manuscript text dump does not include quantitative residual tables for the six cases beyond Δt_f and Δm_f. Adding a compact table of ε_x^∞ and ε_H^∞ at T1–T6 would make the link to Theorem 5.3 explicit.
  4. [§1; References] Reference [18] is listed as “under review.” Clarify what is inherited versus newly specialized (free final time, fixed terminal state, hard bubble BCs, closed-form control) so novelty is self-contained.
  5. [Front matter; §3–§4 headings] Typographical/formatting: TABLE OFCONTENTS spacing; “Boundary-V alue” / “Closed-F orm” hyphenation; occasional missing spaces after commas in math mode. Minor polish only.

Circularity Check

2 steps flagged

Minor self-citation to under-review co-authored framework; residual training and hard-encoded BCs are self-contained, with independent BVP validation (warm-started) showing no definitional circularity.

specific steps
  1. self citation load bearing [Abstract / Sec. 1 / Sec. 4 opening]
    "Building on a recent framework that jointly embeds Pontryagin’s minimum principle and the Hamilton-Jacobi-Bellman equation for general nonlinear optimal control, the proposed OINN approach specializes that idea... Motivated by [18], this section develops the optimality-informed neural network (OINN) approach..."

    The joint PMP–HJB embedding framework that the entire OINN architecture and residual philosophy rest on is imported from an under-review paper co-authored by the present author ([18]). While the free-final-time specialization, hard-constraint constructions, closed-form control substitution, and new Gronwall/stationarity analysis are developed self-containedly here, the foundational premise that residual training on the joint necessary conditions yields a data-efficient optimal policy is justified only by that self-citation.

  2. other [Sec. 6 (Comparison with the Indirect Method)]
    "using the trained network’s own predicted trajectory as the initial guess supplied to the solver. This warm start, rather than a generic flat guess, is what makes the indirect-method shooting problem tractable here: the costates λv, λm are of a substantially different order of magnitude than the states themselves..."

    The claimed independent numerical check (bvp4c solutions at T1–T6) is made tractable only by seeding the solver with the network’s own trajectory and costates. Although bvp4c still enforces the TPBVP to machine-precision residuals and is a different method, the agreement is partially facilitated by the network itself rather than arising from a fully cold-started independent solve; this is a mild validation dependence, not a definitional collapse of the residual claims.

full rationale

The paper's core derivation is self-contained: Pontryagin NCs are re-derived for the free-final-time fixed-terminal lunar problem (Sec. 3), every BC/transversality is hard-encoded by bubble/softplus constructions (eqs. 26–32), the PMP control law is substituted closed-form (eqs. 34–35), and training minimizes only physics residuals (eqs. 36–42) with no labeled optimal trajectories. Scale constants follow dimensional analysis at the envelope center (eqs. 43–49), not data fits. Validation compares the frozen network zero-shot to an independent bvp4c solve of the same TPBVP at six envelope points (plus 80 MC residual draws); reported bvp4c BC residuals are machine-precision (~1e-15 to 1e-24). The sole mild issues are (i) motivation/architecture inheritance from the under-review co-authored [18] and (ii) use of the network trajectory as bvp4c warm-start (necessary for reliable convergence given costate scaling). Neither forces the reported agreement by construction nor collapses the residual-to-error claims (Thm. 5.3, Rem. 5.2) into inputs. No uniqueness theorem, fitted-then-predicted quantities, or renamed empirical patterns appear. Score 2 reflects the self-citation and warm-start dependence without elevating them to load-bearing circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on classical PMP necessary conditions for free-final-time energy-optimal landing, a constant-gravity point-mass lander model, architectural hard constraints, residual training as a proxy for optimality, and several hand-chosen scales/weights/hyperparameters. No new physical entity is postulated; the OINN is a computational construct. Free parameters are conditioning and training knobs, not data-fitted physical constants, but they do affect residual magnitudes and thus the practical error bounds.

free parameters (4)
  • time penalty ρ
    Chosen as 1e8 N² to create a finite free-final-time optimum; directly shapes optimal t_f and the Hamiltonian residual target.
  • loss weights w_V, w_H (=0.01) vs dynamics weights (=1)
    Hand-set relative importance of Bellman and free-time residuals versus dynamics residuals; affects what training prioritizes.
  • adaptive scale constants s_r,s_v,s_m,s_λ,s_V,s_λv,s_tf
    Dimensional-analysis formulas evaluated at envelope center; condition network outputs and nondimensional residuals.
  • network widths (32/64), K=25000, η=2e-3, N=256, h=2e-3, t_min_f=10 s
    Architecture and optimizer hyperparameters that determine achieved residual level and thus claimed accuracy.
axioms (5)
  • standard math Pontryagin minimum principle necessary conditions (costate ODEs, H-minimizing control, free-mass and free-time transversality) characterize the sought extremals.
    §3; training and architecture are built entirely on these necessary conditions, which are not sufficiency guarantees for global optimality.
  • domain assumption Constant lunar gravity, point-mass rocket equation, single throttleable gimbaled engine, no aero forces over the short low-altitude descent.
    §2 dynamics (1)–(3); standard but excludes higher-fidelity gravity, thruster dynamics, and terrain constraints.
  • domain assumption Assumption 1: m(t) ≥ m_dry > 0 throughout training and deployment so residuals stay smooth and Lipschitz constants apply.
    §5; mass bound is not hard-encoded like position/velocity BCs.
  • ad hoc to paper Assumption 3: exact extremal Hamiltonian has a simple transversal zero at the true optimal flight time.
    §5; needed for the first-order flight-time error estimate (55) from residual ε_H.
  • ad hoc to paper Minimizing physics residuals of necessary conditions over domain-randomized collocation yields a policy close enough to the optimal feedback law over Ω.
    Core methodological bet of §4–§5; stationarity of residual loss ≠ global optimality of the control problem.
invented entities (1)
  • OINN architecture for free-final-time lunar landing (condition encoder + main trunk with hard bubble BCs and closed-form PMP control) no independent evidence
    purpose: Represent state/costate/value over Ω while enforcing boundary/transversality by construction and removing control from learning.
    Computational construct specialized from [18]; independent evidence is numerical agreement with bvp4c, not external physical discovery.

pith-pipeline@v1.1.0-grok45 · 27471 in / 3757 out tokens · 37807 ms · 2026-07-12T07:21:09.563887+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Optimality-Informed Neural Networks for Lunar Landing Trajectory Optimization." pith.science (2026). https://pith.science/paper/I3IGE7KN

@misc{pith2026260702741,
  author       = {Pith},
  title        = {Pith review of: Optimality-Informed Neural Networks for Lunar Landing Trajectory Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I3IGE7KN}},
  note         = {Machine review of arXiv:2607.02741}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper develops an Optimality-Informed Neural Network (OINN) approach for the energy-optimal, free-final-time powered descent of a lunar lander from any initial position, velocity, and mass within a bounded operating envelope to a fixed landing site with zero terminal velocity. Building on a recent framework that jointly embeds Pontryagin's minimum principle and the Hamilton-Jacobi-Bellman equation for general nonlinear optimal control, the proposed OINN approach specializes that idea to a lunar landing problem with free time of flight and fixed terminal state. Every boundary and transversality condition is hard-encoded into the network architecture by construction, the closed-form Pontryagin-optimal thrust magnitude and direction law is substituted directly rather than learned, and the remaining state, costate, and an auxiliary value-function output are trained against a physics-residual loss formed entirely from the necessary conditions of optimality, with no precomputed optimal trajectories required. A preliminary theoretical analysis is explored, establishing a stochastic-optimization stationarity guarantee for the offline training procedure, an explicit bound translating the achieved training residual into bounds on touchdown position, touchdown velocity, and flight-time error, and a fixed, input-independent onboard computational and memory cost suitable for real-time deployment. Numerical simulations evaluate the trained policy, with no retraining, against an independently solved indirect-method boundary-value problem at six representative initial states spanning the operating envelope and against eighty additional Monte Carlo simulation runs, demonstrating close agreement with the indirect-method solution and consistently small dynamics and transversality residuals throughout the envelope.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 2 linked inside Pith

  1. [1]

    From new commercial moon landers to as- teroid investigations, expect a slate of exciting space missions in 2025,

    Z. Wang, “From new commercial moon landers to as- teroid investigations, expect a slate of exciting space missions in 2025,” 2025, https://theconversation.co m/from-new-commercial-moon-landers-to-asteroid-i nvestigations-expect-a-slate-of-exciting-space-mission s-in-2025-243645 [Accessed: 2026-06-25]

    2025

  2. [2]

    Landing on the moon is an incredibly difficult feat – 2025 has brought successes and shortfalls for companies and space agencies,

    ——, “Landing on the moon is an incredibly difficult feat – 2025 has brought successes and shortfalls for companies and space agencies,” 2025, https://thecon versation.com/landing-on-the-moon-is-an-incredibl y-difficult-feat-2025-has-brought-successes-and-s hortfalls-for-companies-and-space-agencies-256046 [Accessed: 2026-06-25]

    2025

  3. [3]

    Survey of trajectory opti- mization methods for mars entry and powered descent,

    X. Liu, S. Li, and M. Xin, “Survey of trajectory opti- mization methods for mars entry and powered descent,” Journal of Guidance, Control, and Dynamics, vol. 49, no. 1, pp. 216–239, 2026

    2026

  4. [4]

    Propellant-optimal powered descent guidance,

    P. Lu, “Propellant-optimal powered descent guidance,” Journal of Guidance, Control, and Dynamics, vol. 41, no. 4, pp. 813–826, 2018

    2018

  5. [5]

    A review of pseudospec- tral optimal control: From theory to flight,

    I. M. Ross and M. Karpenko, “A review of pseudospec- tral optimal control: From theory to flight,”Annual Reviews in Control, vol. 36, no. 2, pp. 182–197, 2012

    2012

  6. [6]

    Convex programming approach to powered descent guidance for mars land- ing,

    B. Acikmese and S. R. Ploen, “Convex programming approach to powered descent guidance for mars land- ing,”Journal of Guidance, Control, and Dynamics, vol. 30, no. 5, pp. 1353–1366, 2007

    2007

  7. [7]

    Minimum-landing-error powered-descent guidance for mars landing using convex optimization,

    L. Blackmore, B. Ac ¸ikmes ¸e, and D. P. Scharf, “Minimum-landing-error powered-descent guidance for mars landing using convex optimization,”Journal of Guidance, Control, and Dynamics, vol. 33, no. 4, pp. 1161–1171, 2010

    2010

  8. [8]

    A survey on convex optimization for guid- ance and control of vehicular systems,

    Z. Wang, “A survey on convex optimization for guid- ance and control of vehicular systems,”Annual Reviews in Control, vol. 57, p. 100957, 2024

    2024

  9. [9]

    Real-time optimal control via deep neural networks: Study on landing problems,

    C. S ´anchez-S´anchez and D. Izzo, “Real-time optimal control via deep neural networks: Study on landing problems,”Journal of Guidance, Control, and Dynam- ics, vol. 41, no. 5, pp. 1122–1135, 2018

    2018

  10. [10]

    Real-time control for fuel-optimal moon landing based on an interactive deep reinforcement learning algorithm,

    L. Cheng, Z. Wang, and F. Jiang, “Real-time control for fuel-optimal moon landing based on an interactive deep reinforcement learning algorithm,”Astrodynamics, vol. 3, no. 4, pp. 375–386, 2019

    2019

  11. [11]

    Image-based deep reinforce- ment meta-learning for autonomous lunar landing,

    A. Scorsoglio, A. D’Ambrosio, L. Ghilardi, B. Gaudet, F. Curti, and R. Furfaro, “Image-based deep reinforce- ment meta-learning for autonomous lunar landing,” Journal of Spacecraft and Rockets, vol. 59, no. 1, pp. 17 153–165, 2022

    2022

  12. [12]

    Meta-reinforcement learning guidance, naviga- tion, and control for autonomous lunar landing with safe site selection,

    A. Scorsoglio, B. Gaudet, L. Ghilardi, and R. Fur- faro, “Meta-reinforcement learning guidance, naviga- tion, and control for autonomous lunar landing with safe site selection,”Neural Computing and Applications, vol. 37, no. 22, pp. 17 311–17 340, 2025

    2025

  13. [13]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Jour- nal of Computational physics, vol. 378, pp. 686–707, 2019

    2019

  14. [14]

    Physics-informed neural networks for op- timal planar orbit transfers,

    E. Schiassi, A. D’Ambrosio, K. Drozd, F. Curti, and R. Furfaro, “Physics-informed neural networks for op- timal planar orbit transfers,”Journal of Spacecraft and Rockets, vol. 59, no. 3, pp. 834–849, 2022

    2022

  15. [15]

    Physics-informed pontryagin neural networks for path- constrained optimal control problems,

    A. D’Ambrosio, B. Benedikter, and R. Furfaro, “Physics-informed pontryagin neural networks for path- constrained optimal control problems,”Journal of Guid- ance, Control, and Dynamics, vol. 48, no. 8, pp. 1861– 1877, 2025

    2025

  16. [16]

    Physics-informed deep learn- ing approach to solve optimal control problem,

    K.-M. Na and C.-H. Lee, “Physics-informed deep learn- ing approach to solve optimal control problem,” inAIAA SciTech 2024 F orum, 2024, p. 0945

    2024

  17. [17]

    On nonuniqueness of solutions of hamilton–jacobi–bellman equations,

    A. Misztela, “On nonuniqueness of solutions of hamilton–jacobi–bellman equations,”Applied Mathe- matics & Optimization, vol. 77, no. 3, pp. 599–611, 2018

    2018

  18. [18]

    Unifying optimality principles inspires data-efficient optimal control learning,

    H. Wang, Y . Shi, D. Wu, Z. Wang, T. Zheng, D. Chen, J. Mei, Y . Peng, W. Liu, S. Gong, P. Shi, X. Huang, and L. Cheng, “Unifying optimality principles inspires data-efficient optimal control learning,” under review

  19. [19]

    D. E. Kirk,Optimal control theory: an introduction. Courier Corporation, 2004

    2004

  20. [20]

    Adam: A method for stochas- tic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochas- tic optimization,”arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  21. [21]

    A simple convergence proof of adam and adagrad,

    A. D ´efossez, L. Bottou, F. Bach, and N. Usunier, “A simple convergence proof of adam and adagrad,”arXiv preprint arXiv:2003.02395, 2020

    Pith/arXiv arXiv 2003

  22. [22]

    A new generalization of lemma gronwall-bellman,

    Y . Louartassi, E. H. El Mazoudi, and N. Elalami, “A new generalization of lemma gronwall-bellman,”Ap- plied Mathematical Sciences, vol. 6, no. 13, pp. 621– 628, 2012

    2012

  23. [23]

    Solving boundary value problems for ordinary differ- ential equations in matlab with bvp4c,

    L. F. Shampine, J. Kierzenka, M. W. Reicheltet al., “Solving boundary value problems for ordinary differ- ential equations in matlab with bvp4c,”Tutorial notes, vol. 2000, no. 1–27, p. 4, 2000. BIOGRAPHY[ Zhenbo Wangreceived his B.E. degree in Astronautics from Nanjing Univer- sity of Aeronautics and Astronautics in 2010 and his M.E. degree in Control Eng...

    2000

  24. [24]

    degree in Aeronautics and Astronau- tics from Purdue University and joined the University of Tennessee Knoxville (UTK) as an Assistant Professor

    In 2018, he received his Ph.D. degree in Aeronautics and Astronau- tics from Purdue University and joined the University of Tennessee Knoxville (UTK) as an Assistant Professor. He is now an Associate Professor in the Department of Me- chanical and Aerospace Engineering and the director of the Autonomous Systems Laboratory at UTK. He is a recipient of the ...

    2018