Scenario: Autonomous Driving Agent Modeling in Simulation

"How would you model the behavior of other traffic participants in a driving simulation for training an autonomous agent?"

The Three Main Approaches

APPROACH 1: LOG REPLAY
Record real-world driving data → Replay exact trajectories in sim
Realism: ★★★★★ (it's real data) | Controllability: ★☆☆☆☆ | Reactivity: ☆☆☆☆☆ (agents don't react to ego)

APPROACH 2: RULE-BASED
Define behavior rules: follow lane, maintain gap, yield at intersections
Realism: ★★☆☆☆ (robotic, predictable) | Controllability: ★★★★★ | Reactivity: ★★★☆☆ (reacts but unrealistically)

APPROACH 3: LEARNED POLICIES
Train neural network agents on real driving data (behavior cloning, RL)
Realism: ★★★★☆ (realistic + reactive) | Controllability: ★★☆☆☆ | Reactivity: ★★★★★ (responds naturally to ego)

Detailed Trade-off Analysis

Dimension	Log Replay	Rule-Based	Learned Policies
Realism	Highest — it IS real data. Captures rare/surprising behaviors.	Lowest — agents follow perfect rules. No aggressive drivers, no mistakes.	High — learns distribution of real behaviors including noise.
Reactivity	None — replayed agents follow their recorded path regardless of ego. If ego changes behavior, replayed agents don't adapt.	Reactive but brittle — follows rules but can't handle novel situations.	Best — agents respond realistically to ego's actions.
Controllability	Low — can't make specific scenarios (e.g., "aggressive lane change"). Limited to what was captured.	Full — can script any scenario. "Agent runs red light at T=5s."	Medium — can condition on behavior modes ("aggressive", "timid") but less precise than scripting.
Safety testing	Limited — hard to find edge cases in recorded data. Can't create novel adversarial scenarios.	Best for safety — can script exact failure scenarios. "Pedestrian jaywalks behind bus."	Good — can train adversarial agents that seek to challenge ego.
Distribution shift	None in replay. But ego trained only on replayed scenarios may fail when encountering reactive agents in the real world.	High — rule-based agents behave nothing like real humans. Ego may overfit to predictable agents.	Lowest — if trained on diverse real data, learned agents generalize well.
Compute cost	Minimal (playback)	Low (simple computation)	High (neural network inference per agent per timestep)
Data requirements	Massive logged driving datasets needed	Domain expertise to write rules	Large datasets + training infrastructure

Production Recommendation: Hybrid Approach

Validation Strategy

Statistical validation: Compare distributions of agent behaviors (speed, following distance, lane change frequency) against real-world data. KL-divergence should be below threshold.
Scenario coverage: Maintain a taxonomy of critical scenarios (cut-in, jaywalker, emergency vehicle, etc.). Track which scenarios the simulation covers and at what frequency.
Transfer gap measurement: Deploy ego policy trained in simulation to a real vehicle (closed course first). Measure performance gap between sim and real. If gap is small → sim is "good enough."
Adversarial testing: Train an adversarial agent whose goal is to cause the ego to fail. If the ego passes adversarial testing, it's robust to edge cases.
Human evaluation: Show simulation videos to human drivers. "Does this look like realistic traffic?" Score on 1-5 scale. Target: > 4.0.

Key References

Waymax (Waymo, 2023) — learned agent simulation at scale using real driving logs
SMARTS benchmark — multi-agent driving simulation framework
DriveGAN (NVIDIA) — neural simulation of driving environments