Multi-Platform Drone Swarm with Provable Safety
Same trajectory math on every platform. Provable no-collision. Consistent threat priority across the swarm — without a consensus round.
- 0
- SolvNum collisions across 30 runs
- 1,765
- Float64 collisions across 30 runs
- 100.0%
- SolvNum threat-tier unanimity
- 98.3%
- Float64 threat-tier unanimity
The scenario
Set the picture
A 100-platform autonomous swarm performs distributed ISR and contested-airspace strike rehearsal. The swarm was built up across three procurement cycles, so the same software runs on three slightly different hardware generations. The platforms must execute the same mission plan, never collide (provably, under any input, including degraded sensors and partial-comms), classify and prioritize threats consistently across the swarm, and certify under DoD AI Ethical Principles and emerging ASTM F38 / SAE G-34 frameworks.
This is the Project Replicator profile, the AUKUS Pillar 2 autonomy profile, and the strategic-deterrence profile for the next decade of mass-affordable autonomy.
What it costs today
Swarm certification is the largest unsolved problem in autonomous platforms today. Each hardware generation gets its own certification campaign. A swarm of three generations is three certifications. Certification cost scales linearly (or worse) with platform diversity.
When platforms compute slightly different trajectories from the same plan, doctrine compensates with separation buffers. Bigger buffers mean fewer platforms in the same airspace, which defeats the whole point of mass. Current swarms achieve 'no collisions' through extensive simulation coverage and conservative spacing, not through a formal collision-free guarantee.
Each platform's onboard ML threat classifier produces a slightly different threat list. The swarm 'thinks' different things are the priority threat at the same moment. The certification artifact today is 'we ran 100,000 scenarios and saw no failures' — not the same as 'the system is provably bounded.' Reviewers know the difference.
What changes with SolvNum
Three capabilities, three different parts of the safety case, all from the same data type.
Every platform — regardless of which procurement cycle it came from — computes the same trajectory math, bit-for-bit, attestable by hash. The 'platform variants compute different plans' problem disappears at the math layer.
Per-tick velocity is bounded by scale distance to neighbors. Two platforms k bands apart cannot close that gap faster than 1 band per tick. This is a provable no-collision guarantee at the data-type layer — not the simulation layer.
Threat tier is the scale field of the threat-relevant parameter (closing speed, RCS, time-to-intercept). Every platform that sees the same threat assigns the same tier — by native field equality, not by ML classifier vote. The swarm-wide red list is the union of integer-tagged tracks. Convergence is in zero rounds.
Measurable outcome
What we'll claim — and how it survives review
Each line below maps to a captured number in the demo section. Every number is reproducible from the SolvNum validation suite.
- Provable no-collision swarm convergence — single-tick excursion bounded, inter-platform spacing maintained by data-type property rather than by control-loop heuristic.
- Consistent threat-tier classification across the swarm without consensus protocol overhead.
- Certifiable as a single math model rather than N platform-specific safety cases — order-of-magnitude reduction in certification timeline for hardware-diverse swarms.
- Hardware swap (replacing a procurement-cycle-1 platform with a procurement-cycle-3 platform) requires no re-certification of the math — only of the new platform's host integration.
- Smaller separation buffers achievable because numerical drift between platforms is zero — more platforms per cubic kilometer of airspace.
The demo
What was tested. How. What the script printed.
25-platform simulated swarm, 100 time steps per run, 30 randomized seeds, 2.0 m minimum safety distance, threats spanning 5 orders of magnitude. Both stacks rally to a tight 10 m ring (the 'real' autonomy scenario where collisions actually happen). Two stacks run in parallel: float64 standard formation control + ML threat classifier, and SolvNum stability-bounded formation + scale threat tagging + deterministic trajectory math.
Cross-hardware-variant hash check: same SolvNum scenario run twice on different 'platforms' to confirm the per-tick state hash is bit-stable — a proxy for the cross-hardware identity already demonstrated by the SolvNum cross-platform determinism verification suite.
Live simulation
Animated in-browser simulation of what the demo proves. The numbers underneath are the captured demo output.
Float64 baseline swarm
step 00 collisions
SolvNum bounded swarm
step 00 collisions✓
25 platforms rally to a tight ring (dashed circle). The float swarm converges aggressively, punching through safety bubbles (red collision flashes). The SolvNum swarm slows down as clearance shrinks — it converges asymptotically, never violating the safety distance. Each unique pair collision is counted once. Demo restarts every 100 ticks.
Captured demo output
The numbers the script actually printed.
| Stack | Total collisions | Threat-tier unanimity | Cross-hardware hash |
|---|---|---|---|
| Float64 + ML classifier | 1,765 | 98.3% | — |
| SolvNum (D + B + R) | 0 | 100.0% | f670d10eaec1… (stable) |
Swarm Safety Attestation — Multi-Platform Autonomy
- Platforms
- 25
- Per-tick excursion bound
- ≤ 2.4623× per-step excursion
- Safety distance bound
- ≥ 2.0 m (provable)
- Total collisions (30 runs)
- 0 (vs 1,765 float baseline)
- Threat-tier unanimity
- 100.0% (vs 98.3% baseline)
- Cross-hardware hash
- f670d10eaec1dd3510d7a8d801aeabb100856bd61f07b01d148d624717459d00
- SolvNum table version
- core.K=24, TABLE_BITS=11
Composes with
Where this POC sits in the substrate
Every POC reinforces — and is reinforced by — others. Click through to see how each piece locks into the larger picture.
Mission Rehearsal Parity
Mission Rehearsal Parity is the determinism foundation this POC builds on.
Provable Effector Slew Rate
Effector Slew Rate is the per-effector excursion-limit primitive scaled to swarm-level safety.
Model-Free Anomaly Detection on Sensor Streams
Anomaly Detection uses the same scale-classification primitive on individual sensor streams; this POC uses it for cross-platform threat consistency.
Distributed Sensor Fusion Across a Ship Squadron
Distributed Sensor Fusion is the fleet-internal-fusion variant.
JADC2 Reference Compute Substrate
JADC2 Substrate integrates this with bandwidth (C) for the full reference architecture.
Evidence pointers
Where the claims live in the repo
These are the files a reviewer should run, read, or grep to re-derive every number on this page.
- SolvNum cross-platform determinism verification (x86, ARM, WASM, CUDA)
- SolvNum swarm-formation demo — stability-bounded formation primitive
- SolvNum magnitude-classification demo — scale-based magnitude classification
- SolvNum validation suite — excursion-limit verification
- SolvNum benchmark suite — identity + trajectory + dsbcg verdicts
Previous · POC 08
Counter-Spoofing PNT in GPS-Denied / Jammed Environments
Next · POC 10
Disconnected-Comms Autonomous Handoff
Want to see this in your environment?
Brief us on a program where this POC matters.
ITAR-aware. Air-gapped delivery available. Every claim above traces back to a script in the public repo.