TransforMARS: Multi-Domain Technical Systems

• TransforMARS is a designation for three distinct technical systems in exoplanet simulation, Mars landing guidance, and modular aerial robotics.
• The systems employ specialized methods: transformer-based neural networks for exoplanet synthesis and landing guidance, and graph-based algorithms for fault-tolerant aerial reconfiguration.
• All approaches accelerate domain-specific computations by enabling rapid synthetic modeling, real-time trajectory optimization, and safe reconfiguration under fault conditions.

Searching arXiv for the provided TransforMARS-related papers and uses of the term. TransforMARS is a non-unique designation that appeared in 2025 arXiv literature for three distinct technical systems: a transformer-based generative model for planetary systems trained on the Bern “NGPPS” population synthesis (Alibert et al., 8 Sep 2025), a Transformer-based Successive Convexification framework for six-degree-of-freedom Mars powered descent guidance (Briden et al., 1 Jan 2025), and a fault-tolerant self-reconfiguration framework for Modular Aerial Robot Systems under multiple rotor and unit faults (Huang et al., 17 Sep 2025). The shared label therefore does not denote a single canonical architecture; it denotes separate methods in exoplanet modeling, aerospace guidance, and modular aerial robotics.

1. Terminological scope and disambiguation

In the 2025 literature, the same name is attached to systems with different objectives, state representations, and computational primitives. One use is a decoder-only transformer that models planetary systems as token sequences over a discretized $(\log a,\log M)$ plane. A second use is a transformer-assisted optimal-control pipeline for Mars landing that predicts tight constraints and warm-start trajectories in SCvx. A third use is a reconfiguration framework for modular drones that is formulated through controllability margins, graph connectivity, VMCS construction, and shortest-path planning rather than through a transformer neural architecture (Alibert et al., 8 Sep 2025).

Usage of name	Domain	Primary objective
TransforMARS	Planetary systems	Generate synthetic planetary systems and support conditional prediction
TransforMARS / T-SCvx	6-DoF Mars landing	Predict tight constraints and warm-start reduced SCvx subproblems
TransforMARS	Modular aerial robots	Fault-tolerant self-reconfiguration with continuous in-air stability

A common source of confusion is the assumption that all uses refer to a single model family. The available papers instead describe three distinct technical programs that share only the name and a general concern with structured decision-making under constraints.

2. Planetary-system TransforMARS

In "A transformer-based generative model for planetary systems" (Alibert et al., 8 Sep 2025), TransforMARS is a small decoder-only transformer trained to capture correlations and statistical relationships between planets in the same system. Its stated purpose is to provide large numbers of synthetic planetary systems with little computational cost, relative to direct numerical calculations, and thereby support applications such as guiding observational campaigns and inferring unobserved companions.

The architecture consists of 3 stacked transformer decoder layers with 1 attention head per layer, token embedding dimension $d_{\text{model}}=32$, feed-forward hidden size $d_{ff}=128$ with GeLU activation, vocabulary size $\sim 440$ tokens, and total trainable parameters of $\sim 60\,000$. Tokenization is defined by a uniform $30\times 30$ grid in $\log_{10}(a/\mathrm{AU})\in[-3,3]$ versus $\log_{10}(M/M_\oplus)\in[-4,4]$, with each planet mapped to the rectangle it occupies. A planetary system with $N$ planets is sorted by ascending $a$ and represented as a sequence $d_{\text{model}}=32$0. Self-attention is written as

$d_{\text{model}}=32$1

$d_{\text{model}}=32$2

with $d_{\text{model}}=32$3. The model does not add sinusoidal or learned positional encodings; the ordering by semi-major axis is built into the tokenization.

Training uses 24 365 synthetic planetary systems from the Bern “NGPPS” population synthesis for solar-type stars, with up to 20 planets each and final mass and $d_{\text{model}}=32$4 after 10 Gyr. The reported split is 24 000 systems for training and 1 000 for testing. Optimization uses AdamW with $d_{\text{model}}=32$5, $d_{\text{model}}=32$6, learning rate $d_{\text{model}}=32$7, weight decay $d_{\text{model}}=32$8, batch size 32, and 70 000 gradient steps. The objective is the autoregressive cross-entropy

$d_{\text{model}}=32$9

with best achieved test loss 2.4015, compared to a random-token baseline $d_{ff}=128$0.

Generation proceeds autoregressively. In unconditional sampling, the model starts from an empty prefix, computes $d_{ff}=128$1, samples at temperature $d_{ff}=128$2, and repeats until an end-of-sequence condition such as maximum length 20 is reached. No top-$d_{ff}=128$3 or nucleus filtering was used in the published work. In conditional sampling, observed planets are mapped into tokens and used as a prefix before sampling additional planets. A sampled token identifies one grid rectangle, and $d_{ff}=128$4 and $d_{ff}=128$5 are drawn uniformly within that cell. The decode stage enforces pairwise Hill-stability for adjacent planets:

$d_{ff}=128$6

with

$d_{ff}=128$7

If a sampled pair is unstable, the procedure re-draws within the same rectangle until stability is met.

Validation combines visual, statistical, and machine-learning-driven tests. In the $d_{ff}=128$8-$d_{ff}=128$9 plane, generated and numerical systems show qualitatively similar structures, including peas-in-a-pod trends and spacing. Statistical comparisons cover planet mass, semi-major axis, number of planets per system, Euclidean step lengths $\sim 440$0, and turning angles between consecutive planet segments; agreement is reported as excellent except for minor binning artifacts at small step lengths and angles. In classifier-based tests on 25 000 real and 25 000 generated systems, with 18 classifiers including linear models, SVM, random forests, and feed-forward DNNs, a sanity check on two random halves of the real population yields $\sim 440$1, while real-versus-generated discrimination late in training yields $\sim 440$2–$\sim 440$3, leading to the conclusion that no classifier reliably distinguishes the generated systems from full Bern simulations.

The paper’s use case is the TOI-469 system. Using the observed planet $\sim 440$4 with $\sim 440$5 AU, $\sim 440$6, and $\sim 440$7 m/s, the workflow generates 300 000 full systems unconditionally, selects systems containing a “b-like” planet within $\sim 440$8 in $\sim 440$9 and $\sim 60\,000$0 in $\sim 60\,000$1, leaving $\sim 60\,000$2 survivors, imposes an RV detectability threshold $\sim 60\,000$3 m/s, and then analyzes planets $\sim 60\,000$4. The resulting 2D histogram in $\sim 60\,000$5 versus $\sim 60\,000$6 shows two overdensities coinciding with the actual $\sim 60\,000$7 and $\sim 60\,000$8 planets and a distant giant peak near 1 AU; detectable planets per system range from 2 to 4 with mode $\sim 60\,000$9, and the fraction of systems with a giant planet $30\times 30$0 is $30\times 30$1. The model is provided on the website www.ai4exoplanets.com.

3. TransforMARS as Transformer-based Successive Convexification for Mars landing

In "Tight Constraint Prediction of Six-Degree-of-Freedom Transformer-based Powered Descent Guidance" (Briden et al., 1 Jan 2025), TransforMARS refers to T-SCvx, a Transformer-based Successive Convexification framework for real-time, fuel-optimal 6-DoF Mars powered descent guidance. The optimal-control formulation uses state

$30\times 30$2

with the minimum-fuel objective

$30\times 30$3

The paper specifies rigid-body and translational dynamics, boundary conditions, and nonconvex state and control constraints, including mass lower bounds, glideslope, attitude, angular-rate, thrust-magnitude, and thrust-pointing constraints.

A key ingredient is rotation-invariant data augmentation. Because the Mars landing problem is symmetric about the local “up” axis $30\times 30$4, each feasible trajectory can be rotated by

$30\times 30$5

for $30\times 30$6, while preserving feasibility. This eight-fold augmentation increases a 1,592-sample base set to 11,634 total samples.

The transformer used for tight-constraint prediction receives $30\times 30$7 features and projects them to $30\times 30$8. It uses learned positional encodings, a 4-layer encoder stack, 2 heads per layer, head dimension $30\times 30$9, and dropout $\log_{10}(a/\mathrm{AU})\in[-3,3]$0. The output is a binary vector of length $\log_{10}(a/\mathrm{AU})\in[-3,3]$1 with $\log_{10}(a/\mathrm{AU})\in[-3,3]$2 nodes, so each bit corresponds to one convex inequality at one discretization node. The loss is binary cross-entropy,

$\log_{10}(a/\mathrm{AU})\in[-3,3]$3

An almost identical transformer with $\log_{10}(a/\mathrm{AU})\in[-3,3]$4 and output size $\log_{10}(a/\mathrm{AU})\in[-3,3]$5 predicts the entire trajectory $\log_{10}(a/\mathrm{AU})\in[-3,3]$6 via mean-squared error.

The reduced T-SCvx subproblem is formed by predicting which convex constraints are tight. At iteration $\log_{10}(a/\mathrm{AU})\in[-3,3]$7, the framework computes $\log_{10}(a/\mathrm{AU})\in[-3,3]$8, includes only those convex constraints $\log_{10}(a/\mathrm{AU})\in[-3,3]$9 for which $\log_{10}(M/M_\oplus)\in[-4,4]$0, solves the reduced problem, and updates the trust region and virtual controls using

$\log_{10}(M/M_\oplus)\in[-4,4]$1

This procedure is coupled to a learned solution warm-start, so both the active set and a feasible locally optimal reference trajectory are predicted before the direct optimization step.

Dataset generation uses standard SCvx in Julia plus ECOS on 1,592 random 6-DoF landing instances, with ranges $\log_{10}(M/M_\oplus)\in[-4,4]$2, $\log_{10}(M/M_\oplus)\in[-4,4]$3, $\log_{10}(M/M_\oplus)\in[-4,4]$4, $\log_{10}(M/M_\oplus)\in[-4,4]$5, random attitude $\log_{10}(M/M_\oplus)\in[-4,4]$6, $\log_{10}(M/M_\oplus)\in[-4,4]$7, and $\log_{10}(M/M_\oplus)\in[-4,4]$8. The 11,634 augmented samples are split into 80% training and 20% test, with 3-fold cross-validation and feature standardization to zero mean and unit variance. The transformers are trained in PyTorch with 4,000-step warm-up, two epochs, and batch size 128.

Reported performance includes test binary accuracy 96.45% for the tight-constraint network, train MSE $\log_{10}(M/M_\oplus)\in[-4,4]$9, validation 1.092, and test MSE 1.040 for the solution network. Solve-time comparisons against plain SCvx show mean 14.61 s and median $N$0 s for SCvx at 50 nodes, versus mean 4.95 s and median $N$1 s for T-SCvx, corresponding to a 66% mean reduction and 70% median reduction. On 545 held-out test cases, 100% converged within 20 iterations when warm-started by T-SCvx; in SCvx, a small fraction $N$2 timed out at the maximum iteration count. Final mass is within 0.5% of the SCvx optimum, and all state and control constraints are satisfied to solver tolerances. Inference time is reported as $N$3 ms for the constraint network plus $N$4 ms for the solution network on a modern CPU, and the model footprints are $N$5 MB and $N$6 MB respectively. A space-grade flight processor such as a RAD750 with $N$7 MB SRAM is stated to be able to host these models.

The stated limitations concern out-of-distribution inputs, which may degrade accuracy and motivate OOD detection or expanded training ranges, as well as flight-certification issues such as quantization and code generation. The paper also notes that extension to path-dependent randomized disturbances or mixed-integer constraints may require further architectural or algorithmic enhancements.

4. TransforMARS for modular aerial robot self-reconfiguration

In "TransforMARS: Fault-Tolerant Self-Reconfiguration for Arbitrarily Shaped Modular Aerial Robot Systems" (Huang et al., 17 Sep 2025), TransforMARS denotes a fault-tolerant reconfiguration framework for MARS, where MARS are multiple drone modules physically bound together into a single flying structure. The system consists of $N$8 identical quadrotor-like units rigidly interlocked on a two-dimensional lattice. Reconfiguration is used to mitigate rotor or unit failures while preserving stable flight.

The framework is built around a Controllability Margin (CM) metric. Let $N$9 be the occupied grid cells and $a$0 the faulty units. The adjacency graph $a$1 connects 4-connected lattice neighbors. If $a$2 are the connected components, then each subassembly $a$3 with faulty subset $a$4 has collective control input $a$5, where

$a$6

The weight vector is

$a$7

and the CM is defined as the signed distance from $a$8 to the boundary $a$9:

$d_{\text{model}}=32$00

The global margin is

$d_{\text{model}}=32$01

An assembly is flyable whenever its CM is positive.

The first core routine is VMCS Identification. For each faulty unit $d_{\text{model}}=32$02, the algorithm searches for the smallest rigidly connected subassembly containing $d_{\text{model}}=32$03 that uses only one copy of each module and yields $d_{\text{model}}=32$04. This object is termed the Virtual Minimum Controllable Subassembly (VMCS). The procedure increments the candidate size $d_{\text{model}}=32$05, enumerates connected subsets of size $d_{\text{model}}=32$06 containing the fault, computes their CM, and returns the maximizer once a positive CM is found. The paper notes that worst-case enumeration is combinatorial in $d_{\text{model}}=32$07, although in practice faults are sparse and $d_{\text{model}}=32$08 remains small.

The second core routine is the Sequence Planner. Starting from the initial faulty configuration $d_{\text{model}}=32$09 and a precomputed target $d_{\text{model}}=32$10 that maximizes final CM, it performs three stages. First, it transfers faulty units through VMCS construction; if a faulty unit cannot form its VMCS in the current shape, it solves the mixed-integer problem

$d_{\text{model}}=32$11

to select which normal module to reattach, then generates an $d_{\text{model}}=32$12 trajectory. Second, it performs path clearance by identifying blocker modules that lie on the VMCS trajectories or occupy target cells, and relocating them to waiting positions

$d_{\text{model}}=32$13

via $d_{\text{model}}=32$14. Third, it completes conflict-free assembly of remaining modules by defining target and candidate sets, discarding targets that lie on another target’s path from a virtual start, solving an assignment based on path length $d_{\text{model}}=32$15, and executing the transfers via $d_{\text{model}}=32$16. The overall algorithm performs $d_{\text{model}}=32$17 VMCS constructions and up to $d_{\text{model}}=32$18 relocation steps, each requiring a shortest-path search.

Fault tolerance is enforced through invariance of positive controllability margin. Rotor-level faults are modeled by reducing the per-unit thrust and torque bounds in $d_{\text{model}}=32$19; unit-level faults remove all four rotors. At every intermediate reconfiguration stage, the framework ensures

$d_{\text{model}}=32$20

equivalently

$d_{\text{model}}=32$21

This is the formal condition used to guarantee that the assembly can hover and maintain attitude control continuously during reconfiguration.

Validation is reported in both simulation and hardware. In CoppeliaSim, a $d_{\text{model}}=32$22 grid with two faulty units reconfigures in 4 disassembly/assembly steps with minimum intermediate CM $d_{\text{model}}=32$23, and dynamic tracking errors remain bounded under a robust controller. In a $d_{\text{model}}=32$24 hollow frame with one unit fault, two choices of $d_{\text{model}}=32$25 trade off higher CM against fewer steps. In an 11-unit heart shape with two faults, TransforMARS requires 8 moves with total path length 34 units, whereas disabling path clearance gives 11 moves and total path length 46 units, approximately 35% longer. On a swarm of nine Crazyflie micro-quadcopters, the system executes a reconfiguration from a triangular formation with two immobile faulty units in ten steps, generating collision-free 3D waypoints while maintaining hover stability for the remaining subassemblies.

The paper characterizes the method as the first unified framework supporting multiple rotor and unit failures in arbitrary planar shapes, with integrated path planning and a CM $d_{\text{model}}=32$26 safety guarantee. Its stated limitations are the combinatorial growth of VMCS enumeration and the dependence of real-world reconfiguration on the durability of existing docking and separation hardware.

5. Comparative technical profile

The three TransforMARS systems differ sharply in what is represented, what is optimized, and how correctness is assessed. The planetary-system model represents a system as an autoregressive token sequence over occupied cells of a discretized $d_{\text{model}}=32$27-$d_{\text{model}}=32$28 grid, trains by cross-entropy, and validates by distributional agreement and near-chance real-versus-generated AUC (Alibert et al., 8 Sep 2025). The Mars-guidance system represents discretized optimal-control states and constraints over $d_{\text{model}}=32$29 nodes, trains one network by binary cross-entropy and another by MSE, and validates by solver time, convergence, and trajectory quality against SCvx (Briden et al., 1 Jan 2025). The modular-aerial-robot framework represents assemblies as connected subsets of a lattice graph with faulty subsets, reasons through CM and VMCS, and validates by intermediate controllability, move counts, path lengths, and flight execution in simulator and hardware (Huang et al., 17 Sep 2025).

The first two usages are explicitly transformer-based, but even there the transformer roles differ. The planetary model is a 3-layer decoder-only transformer with 1 head and no positional encodings, used to learn conditional correlations between planets. The Mars-guidance system uses a 4-layer encoder with 2 heads and learned positional encodings to predict active constraints, plus a second transformer for full trajectory prediction. The third usage is organized around graph search, mixed-integer selection, and controllability analysis rather than language-model-style sequence generation. This distinction matters because the shared name can obscure the fact that only two of the three systems are described as transformer neural networks in the supplied literature.

A plausible implication is that the name functions more as a project-specific brand than as a stable methodological category. The underlying methods are not interchangeable: one generates synthetic exoplanetary populations, one accelerates onboard trajectory optimization, and one preserves aerial assembly controllability during reconfiguration.

6. Practical significance, limitations, and outlook

The practical significance of the three systems lies in their domain-specific acceleration of otherwise expensive procedures. In planetary science, the generative TransforMARS is intended to replace repeated direct numerical calculations with a model that can, within seconds, generate millions of statistically faithful systems and answer conditional queries such as predicting planets not yet observed from already observed planets (Alibert et al., 8 Sep 2025). In guidance and control, T-SCvx reduces the effective optimization burden by solving only the predicted tight-constraint subproblem and by warm-starting from a learned reference trajectory, enabling onboard computation of real-time guidance trajectories for a 6-DoF Mars powered landing application (Briden et al., 1 Jan 2025). In modular aerial robotics, TransforMARS converts fault accommodation from a static controllability problem into an explicit sequence of collision-free disassembly and assembly moves that preserve flyability throughout the maneuver (Huang et al., 17 Sep 2025).

The limitations are correspondingly different. The planetary model inherits the discretization and binning artifacts of the token grid and requires rejection-and-redraw within a token cell to enforce pairwise Hill-stability. The Mars-guidance system identifies OOD degradation and flight-certification demands as primary barriers to deployment. The aerial-robot framework is limited by combinatorial VMCS search and by hardware degradation under repeated docking cycles. None of these limitations invalidates the respective methods, but each constrains the regime in which the reported results should be extrapolated.

The publicly stated dissemination channels also reflect their intended use. The planetary model is available to the community through www.ai4exoplanets.com, and the aerial-robot framework provides videos and source code through an anonymous repository. This suggests a trajectory toward reproducible benchmarking and application-specific extension, although the papers themselves describe different future directions: expanded training ranges and certification for T-SCvx, heuristic or incremental VMCS extraction and extension to three-dimensional lattice geometries for modular aerial robots, and continued use of synthetic planetary populations to guide observational campaigns in exoplanet science.