A robotic platform maps outdoor nitrogen dioxide, moves indoors and concludes that an anomaly warrants returning tomorrow. A drone communicates with a ground robot in a dense forest, profiling structure under and above the canopy. An edge device runs simulates multiple scenarios through pre-trained emulators of a regional system. Every component of both exists. The system that ties them together could be next.
Environmental science is being transformed by automated platforms, new low power sensing, digital twins, foundation models and agentic AI. Almost all of it is designed to improve how we measure and simulate the environment, or how we assist a human analyst. A challenging problem is the decision layer: how a system forms a meaningful hypothesis about the environment in front of it, whether real or digital, re-plans what and where to sample or simulate as evidence accumulates, quantifies its own uncertainty well enough to act on it, and knows when to proceed, escalate to a human, or defer. Without all combined, autonomy in the geosciences remains a sophisticated remote control. But when is enough, and who decides? In this session we welcome all contributions on systems that plan, execute or adapt in the geosciences. This can include:
- bespoke and LLM-powered agents with tool use over instruments and models
- software that makes such systems buildable: orchestration frameworks, control interfaces, benchmarks, testbeds and digital twins for training and evaluation
- instrument, sensor and platform development for autonomous operation, with its power, communications and field-hardening constraints
- adaptive, information-driven sampling and autonomous re-tasking of observing networks
machine-generated hypotheses and how they are validated
- closed-loop experimentation controlling instrumentation and iterating on results
- real-time uncertainty quantification and provenance
- accounts of real deployments, including failures
- oversight, escalation protocols, evidential standards and levels-of-autonomy frameworks, including what regulators require before acting on autonomously generated evidence.

Work at any level of maturity is welcome, from deployed systems to position papers, and from all career stages.