Transformer-Informed Trajectory Optimization for Relative Motion in Cislunar Orbits
Walter J. Manuel, Yuji Takubo, Simone D'Amico
math.OC · cs.AI
Quick verdict
Promising extension of ART to cislunar RPOD via dual-arc generation; strong efficiency gains, but single-scenario evaluation and no code release.
Summary
This paper extends the Autonomous Rendezvous Transformer (ART) — a causal transformer trained on solved trajectories to warm-start sequential convex programming (SCP) — to cislunar rendezvous, proximity operations, and docking for the first time. The authors introduce ART-TWIN ('Two-Way INference'), which autoregressively generates two trajectory arcs: one propagated forward from the initial state and one backward from the terminal state, patched at the horizon midpoint. Inspired by forward-backward shooting, this halves each arc's length so error accumulation under chaotic CR3BP dynamics shrinks, converting terminal-state error into a midpoint defect that SCP resolves cheaply. The method is demonstrated on a 6-hour, perilune-initiated rendezvous with a chief on a 9:2 South NRHO, using nonlinear CR3BP LVLH relative dynamics and approach-ellipsoid/keep-out-zone passive safety constraints, trained on 100,000 SCP-solved trajectories. On a 100-case subdataset, ~90% of ART-TWIN warm-starts converge in exactly 3 SCP iterations, whereas the original ART exceeds the 20-iteration cap in over 60% of cases. Across 10,000 held-out samples, ART-TWIN yields a 1.59x faster median SCP runtime than a convex-relaxation warm-start with nearly identical optimality (only ~0.3–0.4 m/s, a 1–2% delta-V difference). Limitations: a single scenario and orbit, no navigation uncertainty or higher-fidelity perturbations, descriptive-only statistics, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- ART-TWIN, via forward-backward dual-arc generation, enables reliable SCP convergence in chaotic cislunar dynamics where the original forward-only ART fails to converge in over 60% of instances.
- This is the first extension of ART beyond Earth orbit and the first use of a nonlinear dynamics model (CR3BP) for ART trajectory generation, demonstrated at perilune.
- ART-TWIN warm-starts yield a 1.59x faster median SCP runtime than convex-relaxation warm-starts across 10,000 test samples, with essentially equivalent optimality (~0.3–0.4 m/s, 1–2% delta-V difference).
- About 90% of ART-TWIN warm-started cases converge in exactly 3 SCP iterations, versus over 60% of ART-0 cases hitting the 20-iteration cap.
Strengths
Well-motivated architecture with a physical mechanism, not a black-box tweak
ART-TWIN's dual-arc generation (Methodology, Figure 5) adapts forward-backward shooting from low-thrust trajectory design [27–31]: halving each arc's length directly targets error accumulation in chaotic CR3BP dynamics, converting terminal-state error into a midpoint defect that SCP cheaply removes. The rationale is transparent and testable.
Clear evidence the architectural change is necessary, not incremental
Figure 8 shows the original ART-0 exceeds the 20-iteration SCP cap in over 60% of the 100 test cases (effectively failing), while ~90% of ART-TWIN warm-starts converge in exactly 3 iterations; ART-0 terminal errors reach beyond 100 km versus ~10 km-order midpoint defects for ART-TWIN.
Large-scale batch evaluation with honest optimality reporting
Challenging scenario selection where linearized models are known to fail
Reproducibility-oriented disclosure
Weaknesses
Single-scenario evidence caps generality
Thin statistical reporting
Missing learning-based baselines
Offline cost and end-to-end timing accounting unquantified
No code, weights, or data release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This AAS conference paper (26-931) extends the Autonomous Rendezvous Transformer (ART) — previously demonstrated only in LEO with linear relative dynamics [10, 24, 25] — to cislunar rendezvous using nonlinear CR3BP dynamics, and introduces ART-TWIN, which generates two arcs (forward from the initial state, backward from the terminal state) patched at the horizon midpoint. Three contributions are stated in the Introduction: the ART-TWIN architecture, the first application of ART beyond Earth orbit and first use of a nonlinear dynamics model, and batch evidence of a 1.59x median runtime speedup over convex-relaxation warm-starts alongside ART-0 failing to converge in over 60% of cases. Claims are scoped to a simulated perilune-initiated rendezvous on a single 9:2 South NRHO segment, and the Conclusion honestly enumerates what is excluded (other orbits, navigation uncertainty, higher-fidelity perturbations).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ART-TWIN and how does it differ from the original ART transformer?
ART-TWIN is an enhanced version of the Autonomous Rendezvous Transformer (ART), a neural network that generates near-optimal spacecraft trajectories to warm-start a convex optimization solver (SCP). While the original ART generates a whole trajectory forward in time from the initial state — accumulating error in chaotic cislunar dynamics — ART-TWIN generates two arcs: one forward from the initial state and one backward from the desired terminal state, joined at the midpoint of the time horizon. Halving each arc's length keeps predictions accurate, and the small mismatch (defect) at the midpoint is easily repaired by the solver.
How much faster is ART-TWIN than a standard convex relaxation warm-start?
In batch tests on 10,000 simulated cislunar rendezvous cases (6-hour, perilune-initiated, chief on a 9:2 South near-rectilinear halo orbit), using ART-TWIN as the warm-start made the SCP solver about 1.59x faster at the median than using a convex-relaxation initial guess, while producing essentially the same fuel-optimal solution (about 0.3–0.4 m/s of delta-V difference, a 1–2% gain). Against the original ART, the contrast is starker: ART-TWIN converged in ~3 iterations for ~90% of a 100-case subset, while the original ART failed to converge in over 60% of cases.
Why does the original ART fail in cislunar space?
Cislunar dynamics (the CR3BP) are chaotic, especially near perilune where the spacecraft moves fastest: small prediction errors compound over the trajectory. In the paper's tests, the original forward-only ART produced terminal-state errors ranging from about 1 km to over 100 km, and the downstream SCP solver hit its 20-iteration cap (effectively failing) in more than 60% of cases. ART-TWIN's two shorter arcs keep errors bounded and move the unavoidable mismatch to the trajectory midpoint, where the solver fixes it cheaply.
Is code or data available for this paper?
The paper does not announce a public code, model, or dataset release. It does disclose the scenario parameters (Tables 1–2), transformer hyperparameters (Table 3), full SCvx* solver settings (Table 4), and the hardware used, but reproducing the results would require reimplementing the entire pipeline, including regenerating the 100,000-trajectory training set by solving the optimal control problems with SCP.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–10 are available only through captions and surrounding prose; exact distribution shapes, outliers, and axis scales in Figures 8 and 10 cannot be verified from the extracted text.
- Equation extraction garbles some sub/superscripts (e.g., midpoint states and Eq. 13 indexing), though the methodology remains unambiguous.
- This is a conference preprint (AAS 26-931, arXiv math.OC, Sep 2026); the compact evaluation scope may partly reflect venue page limits.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.