Concept visualization of a tethered underwater robot approaching purple sea urchins at the edge of a kelp forest

Kelp Recovery · Conservation robotics

Selective robotics for
measurable kelp recovery.

We are developing an operator-supervised ROV system to help restoration teams locate, approach, collect, and document purple sea urchins in bounded restoration areas.

Concept visualization · operator-supervised system

A working software foundation,
ready for physical integration.

Video → observationlocal processing
pipeline operational
Bounded approachprogress checks and
abort logic tested
ArduSub controlvehicle boundary
verified in simulation
Operator authorityoverride and hazard
holds implemented

Targeted action. Measured recovery.

Remove pressure without losing sight of the ecosystem.

When dense purple urchin populations maintain a barren, new kelp can be grazed before a forest re-establishes. Our objective is a selective, traceable intervention that helps qualified restoration teams act precisely in bounded areas.

The ROV is one instrument in a larger system. Each collection can be linked to location, habitat observations, treatment and control areas, and the evidence that matters most: whether a diverse kelp forest persists.

Extend expertise, not replace it

More restoration capacity without abandoning human judgment.

Divers bring ecological awareness and adaptability. Conventional ROVs extend visibility and reach. Our concept combines an operator’s judgment with assisted movement, selective manipulation, and automatic evidence capture.

01

Extend human reach

A tethered vehicle could take on repetitive survey and collection tasks while an operator remains responsible for targets and movement.

02

Pair action with evidence

Candidate sightings, approach decisions, collection events, and locations can become part of one traceable restoration record.

03

Return and measure

Repeatable observations could help restoration teams revisit bounded areas and compare treatment with ecological recovery over time.

04

Preserve selectivity

The design favors supervised, target-by-target work rather than indiscriminate removal or unsupervised collection.

The validation question: Can operator-supervised robotics extend the precision of diver-led restoration into longer, repeatable, data-rich missions?

A supervised restoration loop

See. Decide. Approach.
Verify. Learn.

Autonomy supports the operator; it does not replace them. Every movement passes through explicit authority, proximity, freshness, and abort checks.

  1. 01
    Observe

    Sample the ROV camera stream and vehicle telemetry.

  2. 02
    Identify

    Find target candidates and maintain a stable track.

  3. 03
    Approach

    Issue short, bounded corrections and re-check the scene.

  4. 04
    Confirm

    Stop inside the capture envelope and request operator approval.

  5. 05
    Measure

    Record intervention and ecological follow-up evidence.

03 · Perception pipeline

From underwater video to a target the system can reason about.

We built a local pipeline that turns video into timestamped frames, passes them through an object-detection interface, and converts candidate locations into observations for tracking and navigation.

Video frameCandidateTrackRangeIntent
PERCEPTION / LOCAL
FRAME 0184 · 00:18.4
Illustrative ROV camera scene with six purple sea urchins on a rocky reef target 01candidate target 02candidate target 03candidate target 04candidate target 05candidate target 06candidate
TRACK 03APPROACH CORRIDOR CLEAR
Illustrative interface view showing the target-detection workflow. Generated scene; not field-test imagery.

Prototype evidence

One decision engine.
Several independently testable systems.

01

Video ingestion

Frames in, observations out.

A generated ten-second approach video proved the full media path: encoding, decoding, timestamped sampling, and frame-by-frame inference.

Validated path
encode → decode → sample
Runtime
local Ubuntu processing
02

Decision engine

Move, re-check, or stop.

The vehicle-neutral policy aligns, advances in short pulses, verifies progress, and exits cleanly when it loses the target or fails to get closer.

  • Target lost → abort
  • No forward progress → abort
  • Capture envelope → stop and confirm
03

Control authority

The operator always has the helm.

Assisted motion starts only after an explicit enable and a renewable control lease. Manual override returns authority before the next movement command.

Verified before forward motion
04

Safety supervisor

Hazards veto motion.

Directional proximity sectors represent the contract for rocks, bottom clearance, unknown objects, and kelp or other entanglement hazards.

Rock + kelp hold scenarios passed

05 · Vehicle control

Connected to the real ROV software boundary.

Our adapter communicates over MAVLink with ArduSub—the control software used by BlueROV-class vehicles. In Software-in-the-Loop, it verified heartbeat, mode, arming, telemetry, bounded movement, neutralization, and disarm.

ArduSub 4.7.1official simulator
Command → telemetryintegration path verified
Neutral + disarmverified after every run
kelp-rov / supervised-sitl

01 connect SUBMARINE HEARTBEAT

02 authority OPERATOR ENABLED

03 approach 4 BOUNDED PULSES

04 capture CONFIRMATION REQUIRED

05 cleanup NEUTRAL · DISARMED

PASS / CLEAR CORRIDOR

Local-first, modular by design

An architecture ready to meet a physical vehicle.

Compute-intensive development runs on the project’s Ubuntu workstation. The same interfaces can later connect recorded dives, live camera streams, physical range sensors, and a real ROV without rewriting the mission logic.

Perception

Python · PyTorch · OpenCV · object detection

frames → targets
Decision

Typed state machine · bounded approach policy

targets → intent
Safety

Authority lease · proximity sectors · abort states

intent → permitted intent
Vehicle

ArduSub · MAVLink · pymavlink · SITL

intent → telemetry
Platform

Ubuntu · Docker · project-isolated Kubernetes

local reproducibility

Designed around human judgment

Manual when needed.
Assisted when invited.

01
Manual control

The pilot operates through the standard control interface and owns every maneuver.

explicit enable
02
Assisted approach

The custom policy suggests and executes short, gated movements toward an approved target.

override / fault
03
Hold or return

Any stale signal, hazard, expired lease, or operator input stops the custom controller.

Seeking an evaluation platform

Help build the first physical Kelp Recovery prototype.

WellThought Foundation is seeking a time-limited equipment loan, discounted development system, or technical partnership involving a BlueROV2-class vehicle or manufacturer-recommended equivalent.

Ideal evaluation package

The vehicle and interfaces needed to move from simulation to water.

  • 01
    ROV + topside control

    A complete tethered platform suitable for bench and controlled-tank testing.

  • 02
    Camera and telemetry access

    Live video plus the ArduSub/MAVLink interfaces already exercised by our software.

  • 03
    Manufacturer guidance

    Configuration, safe integration practice, and advice on compatible sensors or manipulators.

  • 04
    Optional additions

    A manipulator, proximity sensing, or ranging system would strengthen later test phases but is not required to begin.

What a hardware partner receives

A documented conservation-robotics integration.

  • Structured compatibility and developer-experience feedback
  • Bench and tank-test results with documented safety behavior
  • Approved photographs and demonstration footage
  • Consideration for a jointly reviewed technical case study
  • Recognition as an early hardware partner, subject to brand approval
Discuss an evaluation platform

From software proof to water evidence

A controlled path from bench to supervised water test.

Completed

Software foundation

Video pipeline, decision engine, simulator control, operator authority, and safety gating.

Phase 01

Bench integration

Verify camera access, telemetry, command arbitration, manual control, neutralization, and emergency-stop behavior.

Phase 02

Controlled tank

Test maneuvering, target tracking, operator takeover, proximity holds, and bounded assisted movement.

Phase 03

Supervised water demonstration

Evaluate representative visibility, range, tether behavior, and documentation with qualified partners and appropriate approvals.

Open the next conversation

Help move a verified simulator stack onto a real evaluation platform.

We are opening conversations with ROV manufacturers, test facilities, restoration practitioners, marine scientists, technical contributors, and mission-aligned funders. The immediate objective is a responsible bench-to-tank program that can inform the first supervised field demonstration.