Autonomy Lab · Mars without round trips

Decide on Mars. Prove it on Earth.

A message to Mars takes up to about 22 minutes one way, and the same back. COSRYX sends the plan and its allowed actions once, as a signed intent. The rover acts on its own inside those signed bounds, and every decision becomes a receipt Earth can check later.

COSRYX-EDGEBuilt · Simulation-validated
COSRYX-MCCBuilt · Simulation-validated
Light-time3–22 min one way
00 The round-trip problem

You cannot drive a rover with a 44-minute lag.

Ask, wait, act, report, wait: on Mars, every loop can cost the better part of an hour.

The one-way light-time between Earth and Mars runs from about 3 to 22 minutes, depending on planet positions. Ground control cannot steer in real time. A rover that asks permission for every step spends most of its day waiting.

Autonomy solves the waiting, but raises a different question: when the rover decides for itself, how does anyone on Earth know what it decided, on what grounds, and whether it stayed within what it was allowed to do?

The Autonomy Lab answers both. COSRYX pre-authorises signed intents: the mission plan and the actions it allows travel once. The rover executes autonomously within those signed bounds. Each decision is receipted on board, chained to the intent it relied on, and checkable on Earth later. No repeated ask-and-wait loops.

The plan travels once. The evidence comes back whole.

01 The timing, drawn out

Ask-and-wait, against one signed intent.

Light-time01 / 06
01 · CLASSIC

Command, wait, report, wait.

Each command crosses the gap; each report crosses back. With 3 to 22 minutes each way, a single step can cost up to about 44 minutes of round trip before the next one starts.

02 · THE COST

Ten steps can take a working day.

At mid-range distance, ten ask-and-wait steps add up to more than four hours of pure waiting. Science time is spent watching a clock.

03 · SIGNED INTENT

One bundle, sent once.

The intent bundle carries the plan, the allowed actions and the bounds: places, slopes, energy, time, instruments. It is signed with Ed25519 and ML-DSA-65 and sent in one trip.

04 · BOUNDED AUTONOMY

The rover decides, and receipts each decision.

Inside the bounds, the rover acts without asking. Every decision is sealed on board as a receipt that names the intent it relied on and the readings it saw.

05 · OUT OF BOUNDS

Outside the bounds, it holds safe and asks.

A slope steeper than allowed, energy lower than allowed: the rover does not improvise. It enters a safe hold, receipts that too, and waits for a new signed intent.

06 · VERIFY

Earth checks what was done, and why.

On the next relay pass the receipts come home. Anyone with the public keys can check that every decision traces to a signed intent and stayed within its bounds.

02 Inside a signed intent

Bounds a reviewer can read.

An intent is not a free hand. It is a short list of what is allowed, signed by the people allowed to allow it. An illustrative example:

# illustrative intent bundle (field names simplified)
intent      "traverse to outcrop B, image, sample if safe"
valid       sol 1,180 → sol 1,183
region      polygon P-17 # four vertices, map frame
bounds      slope ≤ 18° · energy ≥ 40% · drive ≤ 120 m/sol
allowed     drive · image · drill(depth ≤ 5 cm) · reroute(inside region)
on_breach   safe_hold + receipt + ask
approved    ops lead · science lead # two people
signature   Ed25519 + ML-DSA-65 # both required
WHAT

Allowed actions

A closed list. Anything not on it is not allowed, however sensible it looks on board.

WHERE · WHEN

Region and validity

A map region and a time window. Outside either, the intent no longer applies.

LIMITS

Numeric bounds

Slope, energy, distance, depth. Crossing one triggers the breach rule, not improvisation.

03 What it is built from

Products already built and tested on the ground.

ProductRole in autonomyTestsStatus
COSRYX-EDGEOnboard autonomy: executes inside bounds, seals and chains each decision31Built · Simulation-validated
COSRYX-MCCMission control: issues and signs intents, records approvals (needs the V3 AIRGAPKIT, in development)28Built · Simulation-validated
DSRPCarries intents out and receipts home across delayed, interrupted links78 · 100/100 scenariosSimulation-validated
SWARM TRUSTA receipt for every multi-robot decision; tolerates up to ⌊(n−1)/3⌋ faulty units38Built · Simulation-validated
COSRYX-HYBRID-01Quarantines a record when deterministic replay disagrees, before a clean compute receipt is issued24/24Built · definition-of-done sign-off pending
46/0Mars MCC + EDGE conformance checks passed / failed, in simulationSimulation-validated
18/18radiation-integrity checks in simulation (a simulation, not a radiation test)Simulation-validated
3–22minutes of one-way light-time the design assumesSimulated
G3radiation testing is the next gatePending
04 Link loss for months

Store, seal, forward.

If contact with a rover is lost for three or four months, it keeps working within its last valid intent, or holds safe when that intent expires. Either way it keeps recording everything: each record sealed and chained on board, with a sequence number that makes any gap visible.

When contact returns, the rover forwards its backlog to the nearest Mars orbiter it has a relay agreement with; the orbiter sends it on to Earth. Nothing is lost, and nothing can be quietly altered on the way.

LOST

Keep sealing

Hash-chained records on board; no dependence on the link.

EXPIRED

Hold safe

When an intent's validity ends, the rover stops acting and keeps recording.

RETURN

Forward via orbiter

Backlog to the orbiter, orbiter to Earth, every hop receipted by DSRP.

CHECK

No gaps

Earth confirms the sequence is unbroken and both signature legs verify.

05 Many robots

When a swarm decides, the receipt says who agreed.

Future surface operations may use several rovers, drones or landers working together. SWARM TRUST issues a cryptographic receipt for every swarm decision, saying what was decided, when, why, and under whose consensus, even during a communication blackout.

It is Byzantine-tolerant: it survives up to ⌊(n−1)/3⌋ faulty or corrupted robots, scrubs single-event upsets in simulation, and quarantines a rogue unit. Built · Simulation-validated, 38 tests. It rides on COSRYX-NEXUS custody and COSRYX-EDGE.

06 What the lab lets you do

Rehearse autonomy before you trust it.

DRAFT

Write an intent

Draft bounds and allowed actions, and see in simulation where they would hold the rover safe.

Simulated
STRESS

Push it past the edge

Inject steep slopes, low energy and lost links, and confirm every breach ends in a receipted safe hold.

Simulated
REVIEW

Replay a sol

Walk each decision with the copilot, receipt by receipt, and export the evidence for review.

Research preview
07 Questions

About bounded autonomy.

Does the AI decide what the rover may do?

No. People decide, and sign the intent. Onboard autonomy chooses only among actions the intent allows, inside its bounds.

Has this run on Mars?

No. COSRYX-EDGE and COSRYX-MCC are built and simulation-validated on the ground. No flight qualification is claimed, and radiation testing (G3) is the next gate.

What if the signing keys on board are compromised?

Onboard records are sealed with hash commitments and chained; receipts carry two signature legs, and a verifier requires both. Key lifecycle and revocation are part of the trust layer's design.

Could this link Earth, the Moon and Mars?

One verifiable communication and evidence network across Earth, Moon and Mars is a future direction.

08 Pilot

Rehearse one autonomous campaign.

Draft intents, stress the bounds in simulation, and review every decision with its receipts.

Request pilot accessHealth heartbeat