AI in the Glide Phase: What Hypersonic Defense Needs

Most writing on hypersonic defense focuses on the missile. The harder problem is custody and handoff—keeping a maneuvering target in track across sensors and passing a high-quality firing solution to shooters under tight timelines.

In 2025 we saw why this matters: space sensors (HBTSS) helped a Navy destroyer track a maneuvering hypersonic-like target and simulate an SM-6 engagement—the end-to-end "space → ship → shooter" chain actually worked.¹ The lesson for AI teams is clear: the mission wins or loses on continuous track quality, calibrated uncertainty, and edge-constrained inference, not on a single big model.

The mission in 30 seconds

• Glide-phase defense targets hypersonic glide vehicles (HGVs) after boost, while they maneuver at lower altitudes that compress warning and engagement timelines. That low-altitude, maneuvering flight pattern is exactly what breaks traditional radar custody and stresses command timelines.²
• Programs in motion:
  • DARPA Glide Breaker (Phase 2) is quantifying jet-interaction effects between DACS plumes and hypersonic crossflow to inform a future kill vehicle.³
  • Space layer: HBTSS and SDA's Tracking Layer expand global custody and handoff capacity (Tranche 3 proposals in 2025).⁴⁵
  • Interceptor: MDA selected Northrop Grumman to continue the Glide Phase Interceptor (GPI); the U.S.–Japan co-development aims at mid-2030s deployment.⁶

Why “AI for glide phase” is genuinely hard

1. Late detection + aggressive maneuver shrink the window for classification and track-to-engage logic; you need multi-sensor custody with explicit uncertainty, not just a hot classifier.
2. Aero-thermal & plasma effects complicate seekers and comms: strong IR signatures (good) arrive with possible RF attenuation/blackout (bad) as plasma sheaths form at high speeds.⁷ Your models must expect real domain shifts across RF/EO/IR.
3. EW, clutter, and decoys make measurement-to-track association the core problem, not an edge case—multi-hypothesis thinking is baseline.

The AI work that actually moves the needle

1) Custody-first fusion (not model-first everything)

• Objective: maintain continuous track with calibrated error bounds from space sensors to shooters.
• Stack that works: classic Bayesian filters + Multiple Hypothesis Tracking (MHT) for association, with learned residuals to correct systematic error; decision-level fusion that dynamically reweights sensors as conditions drift (e.g., IR upweight when RF is plasma-degraded).

Deliverable an evaluator will recognize: a track quality metric and a track-to-engage latency budget for each handoff (space→C2→shooter), validated against the March 2025 space-to-ship demo timelines.

2) Decisions with uncertainty (abstain > over-confident wrong)

• Hypersonic defense is a shoot/no-shoot/keep-tracking problem under risk. You need uncertainty quantification (UQ) that operators can act on.
• How: pair ensembles or temperature-scaled logits with conformal prediction to produce calibrated sets; add selective (abstaining) classification so the system explicitly withholds advice under low confidence. Measure with risk-coverage curves (AURC) and the newer AUGRC to show safer operation at lower coverages.⁸

Deliverable: plots and thresholds that tie coverage to engagement criteria ("below this confidence, we maintain custody and defer intercept").

3) Edge-constrained inference (D/DIL by design)

• Real nodes are compute-limited and intermittently connected; shipboard decision aids must be CPU-first, bounded-latency, and offline-first.
• Pattern: distilled models (int8/float16) with deterministic fallbacks; hard SLOs (e.g., <50–100 ms for custody updates on active tracks) validated in HIL benches; contention-aware pipelines so triage, not throughput, wins. Use the 2025 HBTSS→Aegis demo as your realism anchor for handoff expectations.

Deliverable: a one-page latency budget from ingest → fusion → UQ → recommendation, showing worst-case CPU execution and queueing under load.

Test like you fight (T&E the KO actually wants)

Move beyond AUC. Field a mission-utility test plan:

• Metrics: time-to-first-track, probability of custody over N seconds, track quality at handoff, false-engagement suppression, plus risk-coverage (AURC/AUGRC) for selective modes.
• Scenarios: maneuvering HGVs with variable altitudes; atmospheric/IR shifts; EW deception; partial sensor outages; space→C2→shooter latency injection reflecting HBTSS/SDA to Aegis paths.
• Artifacts: dataset lineage, eval cards, red-team/adversarial test plans, and a clear abstain logic spec. Map each to the DoD Responsible AI Toolkit so acceptance criteria are obvious.⁹

Architecture, in plain English

1. Space custody: HBTSS + SDA Tracking Layer maintain wide-area detection and tracking of maneuvering threats.
2. Battle management: C2/BMC systems broker handoff with uncertainty attached—your ML must export confidence, not just labels.
3. Shooter aids: near-term SM-6 (demo'd in simulation) and future GPI consume the fused track.¹⁰ Your AI's output is a time-valid firing solution or a principled abstention.

What to publish (so a KO/PM can say "shortlist them")

• Custody-first design doc with the fusion/UQ stack and bounded-latency budgets.
• T&E plan with the mission metrics above and a HIL schedule aligned to known demo chains (HBTSS→Aegis).
• RAI compliance matrix mapped to toolkit artifacts (risk register, data rights stance, audit trails).
• Program literacy one-pager citing Glide Breaker's Phase 2 focus, SDA's Tracking Layer road map, and GPI status (U.S.–Japan).

Bottom line

Glide-phase defense is a custody problem first, an uncertainty problem second, and an edge-engineering problem always. If you build for those realities—before you talk about models—you'll ship systems that survive real timelines, real sensors, and real constraints.

For more on designing AI systems for edge-constrained defense environments, see our post on Edge AI for Defense. Learn more about our Defense & Government AI capabilities.

---

References

---

We build AI systems that keep custody, quantify risk, and run where networks don't. If you're drafting a pre-solicitation around space-to-shooter handoffs or glide-phase T&E, let's talk.

Footnotes

1. MDA and Navy accomplish next step in Hypersonic Missile Defense. DVIDS, March 2025. DVIDS News ↩

2. Congressional Research Service. Hypersonic Weapons: Background and Issues for Congress. Congress.gov ↩

3. DARPA. "Glide Breaker Program Enters New Phase," 2022. DARPA.mil ↩

4. Pentagon launches six satellites to boost missile tracking capability. C4ISRNet, Feb 2024. C4ISRNet ↩

5. SDA Seeks Proposals for Tranche 3 Tracking Layer. SDA.mil, April 2025. SDA.mil ↩

6. Northrop selected to develop anti-hypersonic Glide Phase Interceptor. Breaking Defense, Sept 2024. Breaking Defense ↩

7. NASA. Review of Leading Approaches for Mitigating Hypersonic Vehicle Communications Blackout. NASA Technical Reports Server. NTRS ↩

8. Ding et al. "Revisiting the Evaluation of Uncertainty Estimation and Its Application to Explore Model Complexity-Uncertainty Trade-offs." CVPR 2020. CVF Open Access ↩

9. CDAO Releases Responsible AI (RAI) Toolkit. U.S. Department of Defense. DOD.gov ↩

10. SM-6 Missile Closer To Proving Hypersonic Weapon Intercept Capability. The War Zone. The War Zone ↩