RESEARCH / TECHNICAL REPORTS
← Technical reportsRESEARCH NOTE / 13

Learning-based GNC

Selective Attitude Capture Under Torque Limits

Read the paper ↗

Learn when a faster attitude maneuver is worth it—and when the proven baseline should stay untouched.

AstroUp ResearcharXiv:2610.24763 · 2026Technical analysis
Autonomous decisions

On-orbit servicing windows are short, and every attitude maneuver must finish inside torque and momentum limits with no operator in the loop.

Attitude capture · torque limits · selective control

THE CENTRAL IDEA

Do not replace a qualified attitude controller with a learned one. Train a selector that activates a faster continuation policy only in contexts where paired simulation shows it beats the baseline—and executes the unchanged baseline everywhere else.

Selective Attitude Capture Under Torque Limits

The challenge: fast is good on average, risky per maneuver

Torque-limited attitude capture is the final, decisive part of a slew: the spacecraft must arrive at the target attitude with near-zero body rates while reaction wheels or thrusters approach their authority limits. A more aggressive maneuver profile can shorten the capture and reduce average tracking error across a mission, but braking authority is uncertain—it shifts with inertia properties, momentum state, and actuator health. The same fast profile that helps most maneuvers can push an individual one past its ability to brake.

For a servicer approaching a client satellite, that trade-off cannot be resolved offline once and for all. Ground operators cannot be asked to re-tune capture gains per maneuver, and a conservative universal setting leaves real performance on the table during every short contact window. The paper frames this precisely: improving a controller that is already safe is a different problem from designing a controller from scratch.

A baseline-relative learning problem

The authors treat the existing capture controller as a fixed asset. The learning problem is not "find the best maneuver" but "decide, for this context, whether a faster alternative beats what we already trust." This reframing changes the loss: the metric that matters is the outcome difference between the modified and unchanged controller on the same maneuver, not the absolute tracking error of either. A modification that wins on average but loses on specific contexts is unacceptable for flight, so the optimization target is selective improvement with bounded regret per case.

This baseline-relative view also dictates the data discipline. Contexts used for training, validation, and calibration are kept disjoint, and the evaluation reports per-context outcomes rather than only pooled averages—17 improvements, 31 exact ties, and zero observed regressions for the selective controller. Reporting ties explicitly matters: a selector that cannot beat the baseline should be quiet, not creative.

Three mechanisms of paired-outcome learning

The method combines three mechanisms. First, physically interpretable continuation policies: candidate modifications are alternative speed envelopes for the same capture law, so every candidate remains explainable in terms of braking margin and settling behavior rather than being a free-form network output. Second, separate prediction of task failure and of successful-pair tracking gains: the model estimates both whether a maneuver fails and how much the fast profile helps when both profiles succeed, because these two questions have different costs when wrong. Third, context-dependent rejection: if the predicted advantage is uncertain or absent, the system executes the unchanged baseline.

The rejection option is what makes the scheme flight-compatible. In effect, the learned component is a gate in front of a qualified controller: it can only say "yes, faster here" or "no comment." Ablations reported in the paper show no additional aggregate benefit from branch commitment, active querying, or score subtraction—suggesting the simple gated structure captures most of the available gain.

select(c) = fast, if P(fail∣c, fast) < τ and Δĝ(c) > 0; else baseline

Paired simulation as the label source

Labels come from complete paired simulations: for a given context, both the baseline and the candidate profiles are rolled out under common plant parameters and measurement seeds, so the comparison isolates the controller choice rather than noise. The Basilisk-based campaign uses 108 development contexts, 324 labeling rollouts, and 624 evaluation episodes across 48 held-out test contexts, including disclosed post-freeze diagnostics. Paired labels under shared randomness are what make per-context win/tie/loss accounting credible in the first place.

For teams building similar pipelines, the transferable pattern is the pairing, not the simulator. Whatever high-fidelity environment hosts the attitude loop, labeling should compare candidates under identical disturbances and initial errors; otherwise the selector learns the noise structure of the test rig instead of the physics of the maneuver.

Results: fewer regressions, smaller but safer gains

The selective controller reduces aggregate integrated attitude error by 4.67% relative to the branch baseline, with 17 improved contexts, 31 exact ties, and no observed regressions. The aggressive alternative—always using the fast profile—achieves a 4.81% reduction but worsens 11 cases. The comparison is the paper in one sentence: the extra 0.14 points of average gain cost eleven individual regressions, which is precisely the trade a flight program usually refuses.

The authors are careful about the claim boundary: the evidence supports a restricted empirical trade-off between average improvement and individual regression, and does not establish a closed-loop safety guarantee. For adoption, that honesty is useful—it tells reviewers exactly which assurance gap must be closed by monitoring, envelopes, or formal methods before the gate can be trusted on orbit.

Where the gains appear—and where they do not

Improvements concentrate in principal-axis and near-axis maneuvers, at the braking boundary, and in selected noise/drift cases—contexts where the physics leaves exploitable margin. General coupled maneuvers retain the baseline. This pattern is physically plausible: when rotation axes couple strongly, uncertainty in braking authority grows, and the learned gate correctly stays closed. The map of where gains occur is therefore itself a deliverable: it tells operators which maneuver families are candidates for acceleration and which are not.

A subtle but important detail: 31 exact ties mean the selector frequently faces contexts where the fast profile is neither better nor worse. Treating ties as opportunities would inflate risk without reward; treating them as rejections keeps the system behavior dominated by the qualified controller. This is the right prior for any learning overlay on safety-relevant control.

Engineering mapping and evaluation boundaries

Mapped to an onboard stack, the gate sits between mission planning and the attitude control loop: it reads the maneuver context (target attitude, momentum state, actuator status), evaluates the two predictors, and commits either the baseline or the fast continuation policy before the slew begins. Because candidates are speed envelopes of the same law, switching does not introduce a new control architecture onboard—the verification scope stays close to that of the qualified baseline, and the learned part can be wrapped in runtime monitors that freeze selection if context drifts outside the calibrated domain.

A concrete evaluation plan for adoption: reproduce the paired-label pipeline on the target vehicle model; freeze the selector; then run shadow-mode evaluation across a context matrix that deliberately oversamples braking-boundary and coupled cases, reporting win/tie/loss per context with confidence intervals. Only after the no-regression result survives contexts the training campaign never saw should the gate move from advisory to active. The paper’s disclosed post-freeze diagnostics are a good template for that review.

SOURCE PAPER & REFERENCE MATERIAL

Paired-Outcome Learning for Selective Spacecraft Attitude Capture under Torque Limits

Donghe Chen, Jia Deng, Zehao Jia, Lin Cheng, Shengping Gong

arXiv preprint arXiv:2610.24763 · 2026

Continue to the next report →