Controllable Sim Agents with Behavior Latents

Imagine you're trying to simulate traffic for testing autonomous vehicles. You want the simulated agents to behave like real drivers, but also to be controllable so you can test specific scenarios without real-world risks. Traditional methods often struggle with this because they can’t easily adjust agent behavior or might not respond well to steering commands. This is where things can go wrong, like when agents get stuck in certain behaviors or fail to react appropriately to changes in their environment — this is what's called reward hacking or lack of controllability.

The solution presented in this paper is a new framework called Controllable Neural Variational Agents (CNeVA). It allows agents to learn from past behaviors while also being steered along specific paths. The key innovation here is the use of soft eligibility gates, which help agents respond more smoothly to steering commands instead of getting stuck at rigid thresholds. This means that when you want to test a specific driving scenario, the agents can be guided more effectively without losing their realistic behavior.

Compared to previous models, CNeVA not only maintains a high level of realism but also provides a level of controllability that was lacking before. This is particularly useful for engineers who need to isolate variables and test edge cases safely, making it a practical tool for developing and validating autonomous systems.

The approach introduces a new framework for controllable agents that enhances existing simulation methods.

The results are supported by competitive benchmarks and a clear experimental setup.