DimOS Hackathon · experimental · 2026

PACK MIND

Every team makes robot dogs that share a map.
We made them share a mind.

share meaning, not maps zone names · never coordinates fault-tolerant fleet memory

the problem

Multi-robot usually means merging maps.
That's months of brittle SLAM.

Co-localize everything

Two dogs = two SLAM origins. To share one map you must fuse frames — fragile, and it breaks the moment a loop closes wrong.

Heavy & slow

Map-merge is weeks-to-months of engineering before any coordination happens. Wrong layer to build on in 48 hours.

The reframe

Skip the map. Let each dog keep its own. Share only meaning — what's searched, what's found. Share a mind, not a map.

The whole bet: the useful thing to share between robots is operational memory, not geometry.

the insight

Static zones are a brittle special case of shared memory.

static partition vs shared memory static partition (left) can't adapt · shared memory (right) rebalances

Fixed partition

Alpha→NORTH, Bravo→MIDDLE, Charlie→SOUTH. One finishes early and sits idle. One goes offline and its region is never searched. Indistinguishable from a hardcoded script.

Shared memory

Any body writes what it sees; any body reads and acts. Who-does-what emerges from who's free and what memory says — and it survives a robot dropping.

how it works

Every dog, every tick: sense → remember → choose → plan → move.

pack mind search loop and shared memory one shared map + zone ledger · the A/B knob = share one memory, or each dog keeps its own
SHARED → no two dogs re-search the same ground INDEPENDENT → dogs re-cover each other's trails

what makes it magic

Two beats a script can't fake.

BEAT 1 · HANDOFF

A body writes. Any body acts.

Alpha finds the target and writes it to the pack's memory. I ask Bravo — which was never there — and Bravo walks straight to it. One sees, the whole pack knows.

Bravo acting on Alpha's memory
BEAT 2 · INHERITANCE — the kicker

The mission outlives the robot.

Take the finder offline. Its findings persist and its unsearched ground is reclaimed by the survivor. A database doesn't do that — roles are emergent and fungible.

drop a dog → knowledge loss: 0

Alpha saw it. Bravo remembered it. The pack acted. One memory, many bodies.

the proof · sim A/B, same dogs, same start

Only one variable changes: shared memory.

246 vs 335
ticks to full coverage — shared vs independent. No two dogs search the same ground.
0.80→0.32
Independent: kill a dog and its searched ground vanishes — team knowledge collapses.
drop 0
Shared: kill a dog and its discoveries stay. Knowledge persists; survivors finish.

The honest claim: complete-area coverage, faster & without re-search — and robot-failure tolerance. Not "races to the object faster."

58 tests passing pure-numpy sim · no GPU live + web A/B + DimOS Viewer

why this is worth money

We don't sell robot dogs. We sell the memory layer.

The hardware

commoditizing ↓

Cheaper · longer battery · more agile. Every vendor races on the same axes — a sprint to the bottom. The dog is a replaceable body.

The memory layer — us

compounding ↑

Fleet memory · mission intelligence · shared semantic blackboard. Value grows with fleet size and runtime — hardware-agnostic.

↓ duplicated patrol ↓ labor cost ↑ area coverage ↑ incident response shared audit trail robot-failure tolerance

Run it long enough and the site itself becomes legible — recurring anomalies, chronic blind spots, high-risk hours. Hardware keeps getting cheaper; spatial operating memory only gets more valuable.

where it lands

Pitch with rescue. Land in enterprise patrol.

Phase 1 · Patrol dashboard

Factories · warehouses · campuses · data centers. Shared patrol record, auto incident reports, no re-patrol.

Phase 2 · Mission allocation

Who's nearest? Enough battery? Which zone unchecked? Highest-priority task — assigned straight from shared memory.

Phase 3 · Memory platform

Spatial risk maps · recurring-fault analysis · cross-shift knowledge transfer. The site's long-term operating memory.

honest scope — live = teleop-assisted detection = color filter autonomy shown in sim 1 coordinator → replicate next

the one image to remember

One memory. Many bodies. That's Pack Mind.