Neural operator learning for collision-aware trajectory planning of spacecraft swarms

AI summary

Overview

Research area: Machine learning for astrodynamics — specifically neural operator learning applied to multi-spacecraft trajectory optimization and collision avoidance.

Technical level: Advanced. The paper assumes familiarity with orbital mechanics (Keplerian elements, Gauss Variational Equations), transformer-style attention architectures, and nonlinear least-squares methods (Gauss–Newton, Levenberg–Marquardt, Woodbury identity).

Scope (one sentence): The paper proposes and evaluates a two-stage planner — a self-supervised, permutation-equivariant neural operator followed by a batched Gauss–Newton finish — that produces fuel-efficient, collision-avoiding orbital transfers for spacecraft swarms, trained on ten spacecraft and tested zero-shot on swarms up to 1,000 spacecraft.

What This Paper Is About

Planning orbital transfers for large satellite constellations is expensive because collision-avoidance constraints are pairwise: the number of spacecraft-to-spacecraft and spacecraft-to-debris constraints grows rapidly with swarm size and debris density, so classical optimization must re-solve an increasingly large program for every scenario. The authors instead learn a mapping from distributions of initial spacecraft states, target states, and debris states (plus mission duration) directly to swarm trajectories, then repair each predicted trajectory onto exact two-body dynamics with a cheap per-agent numerical step. The goal is a planner that is fast and scales to swarm sizes where per-agent nonlinear programming is intractable.

Key Contributions

1. A self-supervised neural operator for swarm trajectory planning. It is trained from physics-informed objectives — terminal accuracy, a fuel-effort surrogate derived from the Gauss Variational Equations, closest-point-of-approach (CPA) penalties, and an initial-state anchoring loss — with no optimal-trajectory labels. The operator is permutation-equivariant and time-conditioned, so the same interaction rule applies to any number of sample points and generalizes across swarm sizes without retraining.

2. A batched Gauss–Newton finish enforcing exact orbital dynamics. A single-shooting Gauss–Newton step with Levenberg–Marquardt damping targets each orbit through its conserved specific angular momentum and eccentricity vectors rather than element angles. Using the Woodbury identity, the step collapses to a 6×6 solve whose cost is independent of horizon length and identical for every agent, allowing the whole swarm to be solved as one batched stack of 6×6 systems.

3. An architecture modification for the spacecraft setting. The network builds on prior operator-learning work for mean-field games but replaces a unified attention block with two parallel cross-attention streams — one between agents and target orbits, one between agents and the debris set — followed by a shallow projection head producing per-agent orbital-element predictions.

4. Demonstrated zero-shot extrapolation. Trained on ten spacecraft, the method generalizes without retraining to swarms of 1,000 spacecraft and 11,000 obstacles, and to a full catalog of more than 11,000 objects.

Main Findings

• Zero-shot scaling to large swarms. The operator was trained on missions of 1–12 hours with agent counts N ∈ {1, …, 10} and one adversarial debris object per agent (M = N), yet generalized zero-shot to 1,000 spacecraft and 11,000 obstacles. Performance degraded gradually rather than abruptly out to N = 1000, with terminal error held at 10⁻⁴ to 10⁻² percent throughout.

• Collision avoidance in realistic debris fields. In debris scenarios using real catalogued objects, the operator-warm Gauss–Newton finish (GNw) kept close approaches at or below 0.28 percent of maneuvers within 100 m, while the debris-blind cold finish (GNc) came within 100 m on 2 to 3.5 percent of maneuvers — an order of magnitude more — at a fuel cost within 10 percent of GNw's.

• Adversarial evasion. In adversarial scenarios where one worst-case object is seeded on each method's own debris-unaware path, GNc was struck on 99.3 to 99.8 percent of maneuvers at all sizes, whereas GNw cleared the threat on essentially every maneuver — at most 0.21 percent proximity at N = 1000, and the paper states essentially none of those involves the threat itself — at comparable fuel and accuracy.

• Accuracy measured against an optimal-control reference. At N = 1, IPOPT solved the full nonlinear program attaining 0.0012 percent terminal error at Δv = 0.582 km/s for the minor maneuver and 0.045 percent at 5.52 km/s for the major maneuver. GNw reached comparable terminal accuracy at a fuel cost within 2 percent on the minor maneuver and within 11 percent on the major maneuver, while remaining batched and scalable. IPOPT is reported as tractable only when agent and debris counts are both small; it is described as intractable even at N = 1 against the full catalog.

• Terminal accuracy and fuel across the 2×2 test family. Reported medians over 500 Monte Carlo trials per cell: debris/minor GNw error 0.0007 percent with Δv 0.585–0.590 km/s; debris/major GNw error 0.0186–0.0206 percent with Δv 6.004–6.104 km/s; adversarial/minor GNw error 0.0007 percent with Δv about 0.589 km/s; adversarial/major GNw error 0.0197–0.0213 percent with Δv about 6.06–6.10 km/s.

• Debris input directly shapes the trajectory. Figure 3 shows that when adversarial debris is provided, the predicted trajectory changes to preserve separation above the 100 m threshold, whereas omitting debris information produces multiple predicted close approaches.

• Avoidance as a single capability. The authors report that nominal debris avoidance and defensive evasion are one capability driven by the same conditioning on the surrounding object field, exercised against a deliberately harder threat in the adversarial setting.

Methodology in Plain English

The swarm is treated not as a list of individual spacecraft but as a probability distribution evolving over time. The authors learn an operator that maps the initial spacecraft distribution, the debris distribution, the target distribution, and the mission duration to the swarm's trajectory map, rather than solving a large coupled optimal-control problem for each scenario.

The network is a permutation-equivariant transformer with multi-head attention, so it does not care about the ordering of spacecraft or debris and can accept variable numbers of each. Each agent's trajectory is represented as a smooth deterministic baseline interpolating between initial and target orbital elements, plus a learned residual; the true anomaly is advanced using the physical two-body Keplerian rate plus a learned rate correction.

Training is self-supervised. Four loss terms are combined: a fuel cost estimated by inverting the Gauss Variational Equations with a Moore–Penrose pseudoinverse of the control-affine matrix and integrating squared control magnitude; a CPA penalty using a quadratic hinge, evaluated on a piecewise-linear closest-approach approximation between pairs; a terminal accuracy loss on the first five orbital elements; and an initial-state anchoring loss. A spacecraft–debris safety radius of 1 km applies at all times; a spacecraft–spacecraft radius of 100 m applies only for t ≥ 0.2T with weight w_s = 10, granting a grace window over the deliberately conflicting clustered start while still penalizing unresolved conflict by mid-transfer.

To make avoidance learnable, debris is generated adversarially: for each object the code picks an agent and a random hit time in the latter half of the transfer, rotates that agent's velocity by a random angle in [20°, 75°] about a random perpendicular axis, and offsets the position by a sub-safety-radius near-miss distance of 0.35–0.65 r_d. The authors explain that an exact position–velocity coincidence would put the closest-approach distance at zero, where the CPA penalty's gradient with respect to position vanishes, so the near-miss produces a well-conditioned avoidance gradient instead.

Because the element-space rollout is not the integral of a physical control sequence, a second stage performs a per-agent single-shooting Gauss–Newton step with Levenberg–Marquardt damping over the control sequence only; the state is an exact RK4 rollout from the fixed initial state, so every iterate is dynamically feasible. Residuals are formed on the conserved specific angular momentum h = r × v and eccentricity vector e = (v × h)/μ − r/||r||, which are phase-free and smooth even for near-circular and near-equatorial orbits where element-angle residuals are singular. The 3T-dimensional decision vector is never expanded: the Woodbury identity reduces the step to a 6×6 solve, performed in double precision to absorb the 1/α cancellation when the fuel term is active.

Training details: hidden width 1024, five cross-attention layers, eight heads, approximately 48 million parameters, trained with Adam at learning rate 2×10⁻⁵ under a reduce-on-plateau schedule, gradient-norm clipping at 100, single precision, for 14,000 iterations. Loss weights (λ_f, λ_I, λ_T, λ_S) = (10⁻², 10³, 10², 10²) with hinge scale κ = 10⁴, and anchoring losses normalized element-wise by [1.37, 0.1, π, 2π, 2π]. Scenarios draw N ∈ {1, …, 10} uniformly; a cluster-center orbit is drawn from standard LEO bounds (semi-major axis 6848–8748 km, eccentricity ≤ 0.1, unrestricted angles, periapsis above Earth +100 km); agents sit within a 50 m ball of the center in position and matched velocity, making agent–agent conflict unavoidable at every N ≥ 2. Targets apply one deviation vector at exactly the 1 percent or 10 percent class magnitude to every agent's own initial elements. The horizon is drawn from [1, 12] h. At inference the operator is queried on a uniform grid with a constant 120 s physical timestep in a single batched forward pass.

Evaluation uses a 2×2 family: the maneuver axis sets retargeting magnitude (minor = exactly 1 percent per element, station-keeping; major = exactly 10 percent, rapid response) and the debris axis sets threat construction (ambient catalog objects versus one worst-case object on each method's own debris-unaware path). Proximity is a per-spacecraft rate: the percentage of planned maneuvers passing within 100 m of another agent or debris object. All reported performance values are medians over 500 Monte Carlo trials per cell.

Why This Matters

The paper targets a concrete operational pressure: roughly 16,000 active satellites and an estimated 140 million debris fragments occupy near-Earth space, and between December 2025 and May 2026 alone SpaceX's Starlink constellation executed more than 207,000 automated collision-avoidance maneuvers, over three times its rate a year earlier. Because collision constraints scale pairwise, conventional optimization and rule-based screening both degrade as traffic grows — the former in compute, the latter in conservatism and mutual conflict. This work argues that operator learning grounded in physics can amortize that cost, delivering collision-aware plans in a single forward pass that generalize across swarm sizes and debris densities without retraining.

Real-world applications:

• Constellation-scale maneuver planning. Batch planning of station-keeping and rapid-retargeting maneuvers for swarms of hundreds to thousands of satellites, where per-agent nonlinear programming is intractable.
• Conjunction screening and automated avoidance. The demonstrated adversarial evasion directly models the case of an object on a spacecraft's intended path, the regime that dominates automated collision-avoidance maneuver counts.
• Onboard or ground-segment autonomous guidance. The batched, near-constant-cost finish and single-pass inference suit settings where plans must be recomputed frequently, though the paper does not report onboard hardware benchmarks.
• Mission design and trade studies. Rapid evaluation of transfer fuel cost and safety across many scenario families, using the operator as a surrogate before committing to higher-fidelity optimization.

Industry relevance: satellite constellation operators, space situational awareness and conjunction-assessment providers, and spacecraft autonomy groups all face the scaling problem this paper addresses. The comparison against IPOPT frames the method as a quality-competitive but scalable alternative where a per-agent nonlinear program is no longer practical as a routine online planner.

Future Directions

• Scaling training beyond the tested regime. The operator is trained only on N ≤ 10 agents and horizons of 1–12 hours. Whether training on larger swarms improves the observed gradual degradation out to N = 1000, and how the method behaves on longer transfers, is not established.
• Fidelity of the dynamics model. The method uses two-body Keplerian dynamics — debris are propagated as unpowered objects by advancing only the true anomaly with the other elements fixed, and the Gauss–Newton finish closes onto exact two-body dynamics. Extension to perturbed dynamics such as J2, drag, or maneuver execution error is left open by this formulation.
• Comparison against scalable optimizers at scale. IPOPT is used as a single-agent quality benchmark rather than a scalable baseline, so the paper does not report a head-to-head accuracy or fuel comparison against another method at N = 1000.
• Deployment constraints. Inference is reported on a uniform grid with a constant 120 s timestep in a batched forward pass; the paper does not report runtime, memory footprint, or onboard hardware feasibility, leaving the operational deployment question open. The paper states that Section IV concludes with implications and limitations, but that section's content is not included in the available excerpt.

Target Audience

Researchers and practitioners working at the intersection of machine learning and astrodynamics — particularly those interested in neural operators, permutation-equivariant architectures, self-supervised physics-informed training, and optimal control. It is also relevant to spacecraft guidance, navigation, and control engineers, constellation operations planners, and space situational awareness specialists evaluating scalable collision-avoidance approaches. Readers without a background in orbital mechanics or transformer-based operator learning will find the methodology sections dense.

Authors’ abstract

Autonomous spacecraft swarms must plan fuel-efficient, collision-free maneuvers in increasingly congested orbits, yet classical trajectory optimization scales poorly as pairwise safety constraints multiply with swarm size, and learning-based planners rarely transfer across swarm sizes or debris densities. Here we introduce a permutation-equivariant neural operator that maps distributions of spacecraft, targets and debris to collision-aware trajectories for an entire swarm in a single forward pass, paired with a batched Gauss-Newton finish that enforces exact orbital dynamics. The operator is trained without optimal-trajectory labels, combining self-supervised physics objectives with adversarial threats generated against its own rollouts. Trained on ten spacecraft, it generalizes zero-shot to swarms of 1,000 amid more than 11,000 catalogued objects, matching a per-agent optimal-control solver's accuracy, evading worst-case threats that a debris-blind baseline cannot, and reducing proximity within the swarm several-fold. Physics-grounded operator learning thus offers a fast, scalable alternative to optimal control for crowded orbits.