Where to start

Check our work

If you're investing

← INDUSTRIES

FIELD · LATER

Wide areas, sparse radio, no backhaul for hours.

Farms, mines and construction sites spread robots over distances where one hop is long and a backhaul is occasional. SwUM is built for that: records wait on the robot, and the chain only sees them when a gateway comes within reach.

1,000+

AUTONOMOUS HAUL TRUCKS IN OPERATION WORLDWIDE BY MAY 2022

780

AUTONOMOUS UNITS KOMATSU ALONE HAD DEPLOYED BY 2024

7

SIMULATION SCENARIOS, INCLUDING GATEWAY OUTAGE AND SPLIT-AND-REJOIN

1,000+ · GLOBALDATA, 2022

780 · KOMATSU AT MINEXPO 2024, VIA MINING-TECHNOLOGY.COM

7 · SWUM SPEC V0.2, SECTION 11.6

Long hops, bigger radios, and a chain that can wait.

CLASS M AND L RADIOS

Field work starts when the larger radio classes are proven: LoRa and ESP-NOW frames of 250 bytes, or a Wi-Fi mesh at 1,200. The protocol does not change.

GATEWAY WHEN AVAILABLE

Any robot with backhaul is a gateway. Several may exist at once, and duplicate submissions are harmless by design.

LEARN BETWEEN SHIFTS

Idle time is training time. Outcomes stored on SNIP improve the bidding model and the selector at the next epoch.

THE LOOP

Proof goes to the chain. Data goes to storage.

A finished job becomes a signed attempt record. Records are bundled, and the bundle hash is anchored on SUM Chain when a gateway is in range. The success and failure data goes to SNIP as an episode pack, and a robot with no job trains on it through OmniNode. Nothing is thrown away: even a duplicated job is training data.

Learning loop: prove it on chain, learn from it later When a job is done the robot signs an attempt record; the bundle hash goes to SUM Chain, the raw episode pack goes to SNIP, an idle robot trains on that data through OmniNode, and the new model flows back to the swarm. BUNDLE HASH EPISODE PACK TRAINING DATA NEW MODEL Robot SUM Chain SNIP OmniNode PROOF ON CHAIN · DATA ON SNIP · MODELS BACK

DENSITY

The beacon period adapts to how many robots can hear each other.

With 16 robots in range on a small 802.15.4 radio, beaconing twice a second loses about a fifth of frames to collisions, and once every 1.2 seconds costs a tenth. Proposition 11 gives the number for any deployment, and the communication model learns the period inside a fixed range.

Sparse field sites sit at the easy end of that table.

WHEN

Starts when the bigger radios are proven.

Field deployments wait for the class M and L radios to be proven on the bench and in the factory pilot.

Perception and motion control stay out of scope: the drive controller is a device that accepts velocity or waypoint commands.

WHAT YOU GET

Split and rejoin Scenario S3 stresses partition, duplicate work and convergence after the groups heal.
Gateway outage Scenario S7: records and bundles are stored and forwarded until backhaul returns.
One-way links Scenario S5: mixed radio ranges, with links that work in one direction only.
Density sweep Scenario S6: 8 to 128 robots, airtime limits and beacon period adaptation.