Alex Zongo

Alex Zongo

PhD Candidate, Aerospace Engineering @ GWU · Safe autonomy & multi-agent learning

Research

Data-driven controllers, such as reinforcement learning policies and language models, are increasingly asked to act in dense, safety-critical multi-agent systems. My research asks what it takes to trust them there, especially when the information they act on is degraded, delayed, or adversarial. Air traffic is the proving ground with several interacting agents, hard safety constraints, and sensing that can fail.

Four thrusts, one loop: decide under imperfect information, guard what is deployed, bring in language-level reasoning, and account for what safety costs in operation.

Thrust 1 · Decide

Robust multi-agent RL under degraded sensing

Policies that stay safe when the traffic picture is wrong.

Motivation

Decentralized separation policies rely on positions that every aircraft broadcasts from GPS. Multipath, jamming, or spoofing corrupts those broadcasts, and a policy trained on clean data can act confidently on a traffic picture that does not exist.

Approach

I model observation corruption as a game between the agents and a bounded adversary, derive the worst-case perturbation in closed form, and train policies that are provably close in behavior under clean and corrupted inputs, with no adversarial co-training.

Robust separation under GPS spoofing (schematic)Two aircraft converge on a merge point. Aircraft B's broadcast position is corrupted inside a bounded box; aircraft A, running the robust policy, slows and keeps its separation buffer. merge B A robust policy true position spoofed report bound κ
Schematic. Aircraft A's robust policy keeps its separation buffer while B's broadcast position is spoofed inside a bounded set.

Representative papers

Thrust 2 · Guard

Runtime safety for learned policies

Correct what the policy sees, or constrain what it does.

Motivation

Retraining a deployed policy for every new failure mode is impractical. Safety mechanisms that wrap an existing policy at runtime are attractive, but where they intervene matters.

Approach

I compare runtime filters that share a worst-case state estimate but intervene at different points, on the action or on the observation, and study when each preserves the policy's learned multi-step safety behavior.

Action filtering vs. observation filtering (schematic)A degraded observation feeds a worst-case state estimate. In action filtering, the policy acts on the degraded observation and a barrier check corrects its action. In observation filtering, the estimate is fed to the policy as a corrected input. observation õ (degraded) worst-case estimate ŝ* Action filtering · constrain what it does π(õ) a barrier checkh(ŝ*, a) ≥ 0 ? asafe Observation filtering · correct what it sees π(ŝ*) a
Schematic. Two runtime filters share one worst-case estimate but intervene at different points.

Representative papers

Thrust 3 · Advise

Foundation models in the decision loop

Language models as decision-makers, advisors, and communicators.

Motivation

Large language models can reason about rules and intent, but they are slow, inconsistent, and not grounded in the dynamics of the systems they advise.

Approach

I study how to ground them through simulation-to-language data generation and parameter-efficient fine-tuning (LoRA, GRPO), how to pair them with learned controllers and fallback logic, and how to evaluate them in closed loop rather than on static benchmarks.

A language model in the decision loop (schematic)Traffic from a simulator is described in text, a fine-tuned language model proposes an advisory, and an arbiter passes it to the aircraft or falls back to a learned controller when the model's output is unusable. closed-loop simulation simulation → language AC1: 0.4 nm, closing AC2: crossing, FL004 → action for AC2? fine-tuned LLM LoRA · GRPO arbiter unusable fallback RL policy advisory: AC2 slow down Evaluated in closed loop, not on static benchmarks.
Schematic. Language guidance with an arbiter that falls back to a learned controller when the model output is unusable.

Representative papers

Thrust 4 · Operate

Energy-aware advanced air mobility

What safe separation costs an electric aircraft.

Motivation

Tactical conflict resolution keeps eVTOL aircraft separated, but every maneuver draws on a limited battery, and reserve requirements must account for traffic that cannot be known in advance.

Approach

I couple physics-based eVTOL power models with high-density traffic simulation to quantify the energy cost of deconfliction, and learn estimators with conservative uncertainty bounds for reserve planning.

Energy cost of tactical deconfliction for an eVTOL (schematic)An eVTOL flying between two vertiports deviates around crossing traffic. A bar chart accumulates energy: the deconfliction maneuver adds a small overhead to the nominal energy. V V deconfliction maneuver energy used nominal flown Median overhead < 1.5% · 95th percentile 3.8–5.3% ~72k en-route sections, 10–60 aircraft (ICNS 2026)
Schematic. A deconfliction maneuver adds a small energy overhead; statistics from the ICNS 2026 study.

Representative papers

Earlier work

Reinforcement learning for fault-tolerant attitude control of fixed-wing aircraft (master's research, Tsinghua University).

Supported by NASA through the University Leadership Initiative and by the Office of Naval Research. All publications