ASTRA: A Scalable Next-Generation ATCO Training Simulator with Autonomous Simpilots

Synthesis: ASTRA uses autonomous AI sim-pilots to deliver scalable air traffic control (ATCO) training, reducing reliance on human role-players while maintaining realistic scenario complexity. By replacing specialized human trainers ("simpilots") who must role-play both pilots and ATCOs, the system increases training capacity and uses LLM-driven simulation to support professional training at scale.

ASTRA uses autonomous AI sim-pilots for scalable air traffic control training, reducing dependency on human role-players while maintaining realistic scenario complexity. Air Traffic Control Operators (ATCOs) are vital to the safe, orderly, and efficient flow of air traffic, yet training capacity is constrained by reliance on specialized human trainers known as simpilots, who must role-play both pilots and ATCOs in a simulated airspace. ASTRA's autonomous sim-pilots remove this bottleneck, enabling scalable adaptive practice in a simulated environment, with implications for AI tutoring and training-system design.

What this means for practice

• Designers. Match the speech models to the accent and terminology of the trainees before promising scalability: the fine-tuned ASTRA pipeline reached a word error rate of 23.45% on Singaporean-accented aviation speech, where off-the-shelf systems reached 107.80%.
• Designers. Split radiotelephony assessment into a deterministic rule layer and an LLM layer: ASTRA validates phraseology and separation events with hard rules and scores meaning with BERT-based semantic similarity, which keeps scoring stable while tolerating paraphrase.
• Software developers. Budget prompt compilation as real engineering work rather than an afterthought: DSPy optimization raised the accuracy evaluator from 83.8% to 91.7% and lifted brevity and completeness to 89.7% and 88.1%.
• Software developers. Translate the human role-player's role into authorable scenario artifacts — aircraft profiles, triggers, node-based event sequencing — so instructors can build a training situation without scripting every utterance.
• Researchers. Treat an automated LLM evaluator as a secondary signal: ASTRA's LLM judge diverged from human raters on prosodic and acoustic quality, so calibration against a larger human panel is required before high-stakes scoring.

Limitations

• Automatic speech recognition was fine-tuned exclusively on Singapore ATC data, so generalization to other airspaces, accents, and languages is untested; error rates rose on the out-of-domain ATCOSIM and MNSC datasets.
• The human mean-opinion-score study of text-to-speech used 21 raters evaluating 9 clips each, a sample the authors state limits the statistical robustness of the results.
• The simulator supports only predefined scenarios; instructor live-editing of scenarios and adaptive scenario generation are described as future work rather than evaluated.
• All reported evaluations are component-level (word error rate, mean opinion score, evaluator alignment) — no trainee cohort was run, so gains in learning or transfer, and comparison against a human simpilot baseline, remain open.

Citation

Ethan Chew, Enjia Wu, Iruss Eng Wei Yeow, Ian Weiqin Lim, Ranen Sim, Brandon Koh Ziheng, Kaleb Nim, Caden Toh Jun Yi, Wei Dong Soin, Darius Kai Keat Koh, Galen King Yu Tay, Prannaya Gupta, Jonathan Ee Fang Koong, Yong Zhi Lim (2026). ASTRA: A Scalable Next-Generation ATCO Training Simulator with Autonomous Simpilots.