Research
Self-correcting models of the real world, built and proven at sea.
One idea, worked at every layer
Our thesis: genuine autonomy in messy environments does not come from bigger datasets. It comes from agents that carry an internal model of their world, test it against every sensation, and never stop correcting it. Everything on this page is that one idea, run outside the laboratory.
A camera watching the shore measures its own motion; that steadies the inertial unit, and the inertial unit in turn steadies the image. Each sensor sharpens the others.
Programmes in production
Profundi — a reality compiler for the sea
Given a region, Profundi gathers every available layer of truth about it — hydrographic soundings, electronic charts, elevation models, satellite imagery, marine aids — and fuses eight unreliable, scattered sources into one continuous seafloor model with per-point confidence. Where sources agree the surface is smooth; where data is dense it is faithful; and where evidence runs out it says unknown rather than manufacturing confident terrain. Profundi runs today on the vessel’s edge computer.
Vessel awareness model — real-time state estimation on an instrumented sailing vessel
A live physical model of a sailing vessel — at anchor and under way — fused with her onboard sensors in real time. It ingests a 50 Hz stream of inertial and magnetometer data, satellite navigation, AIS, and environmental sensors, and maintains a continuous estimate of the vessel’s attitude, position, and motion. Its sensor calibration was fitted from 3.96 million real at-sea samples, resolving to a 0.33° residual — measured from the sea, not assumed from a datasheet.
A supervisory agent — months in production over live infrastructure
The group’s supervisory agent has run unattended for months over our own compute infrastructure — watching its telemetry, deciding what is normal, and investigating before it alarms: faced with a surprise, it gathers fresh evidence and re-evaluates, so transient blips resolve silently. Every judgement is recorded. To date it has produced more than five hundred dated run reports — belief updates, prediction errors, recoveries, and self-classified false alarms: a logged operating history of an autonomous system reasoning about itself.
A working vessel, not a lab rig
The group’s proving ground is not a bench but a working vessel. SV Endurance, a ten-tonne Camper Nicholson 35, operates off-grid in the Salish Sea — solar-powered, electrically propelled, self-sufficient. Over thousands of tracked nautical miles her instruments generate the continuous, real-world streams our models learn from, logged under way in fog, chop, and sensor dropout. She is at once the platform, the power plant, and the primary source of the group’s evidence.
The broader programme
Continuous world models
Representing the environment as continuous fields rather than brittle meshes, so that noisy, partial, multi-source evidence combines into one queryable model of reality.
Predictive simulation
An agent’s mind’s eye: simulating what each sensor should report moments ahead, so reality can be compared against expectation as it arrives.
The epistemology of machine perception
Sensors disagree. We treat conflict between partial models as information, and require systems to distinguish “the world changed” from “my model is wrong”.
Auditable autonomy
Append-only, tamper-evident records of what was sensed, believed, and decided — autonomy whose every disposition can be inspected, replayed, and ultimately proven.