| name | autowebworld-synthetic-web-environments |
| title | AutoWebWorld: Synthesizing Infinite Verifiable Web Environments |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2602.14296 |
| keywords | ["web automation","agent training","synthetic data generation","finite state machines","trajectory collection"] |
| description | Generate synthetic web environments at scale by specifying websites as Finite State Machines with explicit state transitions, then programmatically executing GUI actions to collect verified interaction trajectories. Reduces trajectory cost from $0.15–$1.00 to $0.04 per sample while generating 11,000+ verified trajectories with deterministic, executable validation requiring no external judges. |
AutoWebWorld: Deterministic State-Machine Web Environments for Agent Training
Training web automation agents requires diverse, labeled interaction trajectories, but collecting real-world web data is expensive (requiring human judges or reward models), slow (real websites have variable latency), and non-reproducible (website layouts change constantly). Existing synthetic approaches either produce toy websites or generate trajectories without programmatic verification, limiting scalability and trustworthiness.
The core insight is treating websites as Finite State Machines where all state transitions are explicit and deterministic, enabling programmatic trajectory verification without external judges. By making state transitions machine-readable, the system can enumerate valid action sequences through breadth-first search and validate them through execution.
Core Concept
AutoWebWorld uses a four-stage pipeline grounded in FSM theory:
- FSM Specification: Multi-agent framework generates a formal FSM for each website theme, defining pages, allowable states, deterministic preconditions, and state transition effects
- Code Generation: Coding agents translate the FSM into executable Vue.js front-end code, iterating until the project builds successfully
- Trajectory Enumeration: Breadth-first search explores the FSM graph to generate all valid action sequences, expanded into atomic GUI operations
- Execution & Filtering: Execute trajectories on synthesized websites using Playwright, retaining only those where all steps succeed
The key advantage: unlike real websites with implicit state, AutoWebWorld defines states explicitly, enabling deterministic validation without human review or learned judges.
Architecture Overview
- FSM Graph: Nodes represent website states (pages, form states); edges represent actions (clicks, form fills) with preconditions and effects
- Web Generation Pipeline: Convert FSM to Vue component hierarchy with routing, form handling, and state management
- Action Executor: Atomic operations for GUI interaction (locate element, click, enter text, submit form)
- Trajectory Validator: Execute full trajectories and mark as valid only if all steps complete without error
- Cost Evaluator: Track generation cost (LLM calls), execution cost (Playwright runtime), and cost per verified trajectory
Implementation
Define an FSM specification as a structured schema, then generate websites programmatically:
def ():
spec = {
: theme,
: [
{: , : [, , ]},
{: , : [, , ]},
{: , : [, , ]}
],
: ,
: [
{
: ,
: ,
: ,
: ,
: []
},
{
: ,
: ,
: ,
: ,
: []
}
]
}
spec
