DREAMS Laboratory

Why environmental science first

Reefs, fault scarps, drylands, and planetary surfaces are where autonomous science can be developed responsibly. When a twin is wrong here, the cost is a bad map — a survey to repeat, a sample to retake — not a disaster. That makes these domains the proving ground: demanding enough to need real autonomy, forgiving enough to learn from failure in the open.

On autonomous science and AI

The lab's long-term aim is autonomous science: systems that do not only choose the next measurement, but propose the questions worth measuring. We pursue it in stages, and in the open. Today our systems decide where to look inside questions people pose; each step toward choosing the questions is earned by calibration against the physical world, in domains where being wrong costs a map rather than a life. A machine scientist will not be free of bias — its priors and objectives are choices too — but its biases can be written down, audited, and tested, which is more than most of us manage for our own.

Terrestrial

Projects span geomorphology, drylands ecology, and precision agriculture. Systems and algorithms operate as a network of autonomous ground and aerial vehicles, and ground-based static sensors, enabled by cyber-physical twins.

Terrestrial Robotic Observing Network (TRON)

QuAD-FLIP: Quadrotor Aerial Delivery with Flipping for Limited-access Inspection and Probing

EarthRover: Affordable and Sustainable Mobility Autonomy for 4D Environmental Monitoring

Aquatic

Robotic boats and autonomous underwater vehicles for limnological, littoral, and coastal mapping and sampling of the water column and the benthos.

Collaborative Robotic Aquatic Laboratory (CoRAL)

Cyber-Physical Twins

Digital and physical twins leveraging cyber-physical systems, for simulation studies, dynamical analysis, and predictive modelling of environmental processes.

DeepGIS decision support system

deepgis.org · deepgis-xr on GitHub

Virtual Shake Robot and ShakeBot

OpenUAV Simulation Testbed

Planetary

Surface feature mapping, vision systems, and digital twins, with a focus on Mars and the Moon.

Sub-centimeter Resolution Lunar Digital Twin for Collaborative and Immersive Mission Planning (NASA STTR)

High-altitude balloon payload with attitude control and predictive scheduling

Theory

Distinction dynamics is the question beneath the hardware made precise: what an agent can tell apart is not fixed, and every observation changes it. Spectral kernel dynamics treats that capacity — the kernel — as a dynamical variable governed by Maximum Caliber, yielding the most probable paths a perceiving system takes through what it can represent and a lower bound on what that change costs.

Spectral kernel dynamics via Maximum Caliber · applied to planetary surface graphs

Craft and public art

The twin pipeline also runs the other way. In Engineering Mythology, DeepGIS served as a 1:1 template engine for master artisans in Odisha, India, building Navagunjara Reborn, an 18-foot sculpture for Burning Man, produced through Earth Innovation Hub. The artisans work as knowledge-peers rather than fabricators, with craft mastery and engineering treated as interacting knowledge systems. The sculpture's 2026 reincarnation tests the same claim a digital twin makes: that identity can persist through documented re-instantiation.

Twin-to-reality transfer

Every twin on this page is useful only if inference made on it transfers to the original — and the case that matters most is the hardest one: novel regimes the twin was never calibrated in. A twin can be precise and still wrong. We treat calibration of twin against world as the lab's central unsolved problem, not a solved preprocessing step.

Precariously balanced rocks are the flagship test. Their fragility is physically measurable, so predictions from the Virtual Shake Robot can be held against ShakeBot on the bench, where overturning is observed directly rather than inferred · calibration, benchmarking, and sensitivity (2026)