# Autonomous systems learn through simulation and reward

Markdown copy of https://realalpha.app/s/autonomous-systems-learn-through-simulation-and-reward/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Brandon Tseng](https://realalpha.app/builders/brandon-tseng/) · Brandon Tseng is president and cofounder of Shield AI. · Co-founded Shield AI, maker of autonomous V-BAT drones and AI pilots flying F-16s; former Navy SEAL.
- Host: [Shawn Ryan](https://realalpha.app/people/shawn-ryan/) · Host, The Shawn Ryan Show
- Show: The Shawn Ryan Show
- Episode: [Brandon Tseng – Shield AI’s X-BAT: The First AI Fighter Jet to Outsmart Top Gun | SRS #247](https://realalpha.app/v/p8DjGa16PQM/), Oct 23, 2025
- Moment: 1:15:48 to 1:20:50 of the episode (5 min), playing on https://realalpha.app/s/autonomous-systems-learn-through-simulation-and-reward/; the episode from there: https://realalpha.app/v/p8DjGa16PQM/?t=4548

## Quote

> “you basically give it a goal and then behaviors that you believe are positive, **you reinforce those behaviors**, you reward those behaviors.”
> — [Brandon Tseng](https://realalpha.app/builders/brandon-tseng/) at [1:19:36](https://realalpha.app/v/p8DjGa16PQM/?t=4776) · The Shawn Ryan Show

## Takeaways

1. An autonomy stack combines mapping, controls, reasoning and planning.
2. Reinforcement learning rewards desired behaviors and discourages others.
3. Simulation results are checked against real-world performance.

## About this segment

Brandon Tseng breaks down the software modules in an AI pilot, then explains how reinforcement learning trains systems across millions of simulations. The method underpins tasks from room-clearing to dogfighting.

From “Brandon Tseng – Shield AI’s X-BAT: The First AI Fighter Jet to Outsmart Top Gun | SRS #247”, Oct 23, 2025. [Watch the full episode](https://realalpha.app/v/p8DjGa16PQM/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[1:16:00](https://realalpha.app/v/p8DjGa16PQM/?t=4560) You just called it the expat. I mean, you know, is it just you switch it on and point it in the right direction? Yeah, so for the quadcopter, there's a button on the side, you press it, like spin its rotors and like take off from your hand and like go inside the buildings. The way that like it works from a technical perspective and like this is a simplification of it, but there's a series of software modules and every self-driving car or humanoid robot is working like this um it and like every one of shield ai's products is working like this the software modules that make up this ai pilot or this autonomy stack uh you have a state estimation module state estimation is like where in the world am i you have a mapping module it builds a map uh you have a controls module you have uh right that's like how to actuate motors

[1:16:52](https://realalpha.app/v/p8DjGa16PQM/?t=4612) um yeah you know your your steering mechanisms um you have um a reasoning model what we call global reasoning you can think of this as like the mission is the mission to go inside a building and clear the threats is the mission to dogfight f-16s is the mission to find surface to air missile systems like what the mission is that's what the reasoning module is then you have like path planning module which is like okay what are the steps that i'm going to take to execute uh said mission so those are the different software modules that make up an autonomy stack The way that, you know, it operates without GPS, right, it's building a map of the room, its environment, whether it's using cameras, whether it's using LIDAR, whether it's using radars, to basically build a map of its environment, and then it's estimating its position

[1:17:46](https://realalpha.app/v/p8DjGa16PQM/?t=4666) based on the map that it's built. Again, I think a lot of people try to, like, over-complicate autonomous systems. At the end of the day, like, you and I, we are autonomous systems, and we operate the same way these robots do in the sense that we sense the world with our, you know, five senses, then we think about the world using our brain and then we actuate our muscles supported by our skeleton to maneuver around the world. When we come into this room, like we don't think about it this way, but like we've made a map of this room in our head and we can now navigate off of this map. I know that wall is, you know, 10 feet away. This wall is eight feet away, right? You're six feet from me. These These autonomous systems are doing the exact same thing.

[1:18:29](https://realalpha.app/v/p8DjGa16PQM/?t=4709) And just like we learn from our experiences, these things are learning from their experiences. How do they learn from their experiences? Yeah. So, one of the principal ways that we developed a methodology called reinforcement learning. And so, this was pioneered... There's probably academic papers, I would say. like it became pretty famous with like alpha go alpha star, um, where it was, um, they would put these, uh, alpha go, you know, beat a world champion, uh, at the game of go, right. It's any time reinforcement learning is really an, a good learning methodology when the number of variables or outcomes are massive, right. In the, the multiple, like trillions of outcomes, right then which i i couldn't bring up what the outcomes of the game of go are but right it's you know it's almost uncountable yeah right same thing with like the game of starcraft which was

[1:19:29](https://realalpha.app/v/p8DjGa16PQM/?t=4769) alpha star it's like the number of like moves are uncountable in terms of like the number of scenarios that's where you find something like reinforcement learning really shine right in the real world the number of moves that you have are uncountable um and so uh what you're doing is you're taking you know, you'll, you basically give it a goal and then behaviors that you believe are positive, you reinforce those behaviors, you reward those behaviors. And then, um, things that are negative, like you don't reward those behaviors. And what you'll do is spin up millions of simulations where these things are just playing against each other millions of different times and they just continue to improve. And then you have your engineers, right? They architect this, this This training environment, this, you know, this, I think we've called it a dojo before.

[1:20:19](https://realalpha.app/v/p8DjGa16PQM/?t=4819) But they just keep learning from those experiences. You check in, you validate, you cross-validate, you make sure it works in the real world. And that's how you're teaching these systems anything from clearing a room to dogfighting to, you know, hunting surface-to-air missile systems to navigating, you know, yeah, any unmanned system. Wow. Wow.