Research at ASTRONOM
We publish open research on agentic web data, simulation augmentation, and Vision-Language-Action models for autonomous browser agents.
Why this matters
The bottleneck for autonomous AI agents in 2026 isn't compute or model architecture — it's real-world web behavior data. Closed labs hoard internal tool-use traces. Public datasets are stale, narrow, and over-fit. Without diverse, consented, real human web behavior, agents fail outside the demo.
ASTRONOM treats web behavior as the new training fuel — and treats contributors as owners, not raw material.
What we're working on
Generating thousands of robust training samples from a single human demonstration via DOM perturbation and synthetic rendering.
Differential privacy and selective redaction of personal data in captured browser trajectories.
Compressing large foundation agent models into deployable browser-runnable policies.
Coming soon
The first technical report — "Browser-Native Trajectory Capture for Agent Training" — will be published Q3 2026 alongside the network mainnet launch. Subscribe via early access to be notified.