Building worlds where humans, robots and AI agents learn, play, and evolve together.
I spent many years making AAA games, then moved to simulation for defense and robotics. My environments train and test robots and AI agents before they act in the real world, so they protect people and stay aligned with the humans they serve.
Ten projects with public code, live demos or video. Each has a project page or an engineering journal with the problem, the design and the results.
Autonomous operations on infrastructure you own. A self-hosted autonomy harness for physical AI: agents rehearse in simulation, command robots in the field through the same controls, and build their own connectors to new sensors, robots and systems.
An operator typed “fly to Times Square”, and the pilot agent planned the route and flew it, using only the vehicle it was allowed to command.
A Rust and CUDA runtime for JEPA world models. SkyJEPA, rebuilt from its paper, learns how a drone moves and plans its rotor commands, and the JEPA Gym compares model variants on the same drone in MuJoCo.
The learned model made a hand-written flight controller more accurate, and it passed every tracking run across three training seeds and four test conditions.
Experiments with world models and 3D reconstruction: turning video into maps, generating worlds for robots, and opening each pipeline up stage by stage.
Six seconds of generated robot video became a camera path and a full 3D map in about 20 seconds on one GPU, and every stage replays in the browser.
Robots built, broken and measured in simulation: a robot dog that keeps walking on three legs, an amphibious quadruped, rovers on a failing radio network, wind-aware drone delivery, a surgical robot and a two-armed hexapod.
Every project has a journal that shows how it was built, with its videos, run instructions and measured results.
Language and embedding models playing Doom: a trained ModernBERT classifier that reads the game state as text, GLiNER2.5 zero-shot and then fine-tuned, and EmbeddingGemma 2 reading the screen.
EmbeddingGemma 2 plays from the screen with no training: each decision compares crops of the frame against text prompts.
I developed the simulator shown in Shield AI’s Hivemind fixed-wing demo. Hivemind is Shield AI’s autonomy software for military and commercial aircraft.
Hivemind flies the aircraft. The world it flies in is my simulator.
An execution model for simulated worlds. A Lean 4 kernel checks every state change, and each run can be replayed. Agents, people and planners propose actions; the kernel accepts or rejects each one.
Play a mission in the browser and see why every order is accepted or refused, then rewind and try another path.
A Rust engine that models a problem as accounts, assets and exchanges. Search, optimization and reinforcement learning propose exchanges; the engine validates each one and keeps a replayable trace.
The same engine solves a maze, runs a market and schedules a league of agents, and shows the reason behind every decision.
A reinforcement learning and game development harness. Coding, design and NPC agents use it as a world model: they explore procedurally generated 3D worlds in simulation and train policies inside the development loop.
A whole fleet retrains in under 20 minutes on one GPU, so agents can test a design change in the same session.
An open-source simulation engine and AI platform for training perception models on synthetic data and testing robots in simulation.
Robots and perception models learned from synthetic worlds before they met the real one.
Short videos of current experiments. Hover over a card to preview it; select it to watch.
Six seconds of AI-generated video from a robot rolling through a junkyard, turned into the camera’s path and a 3D map of the scene.
A learned policy steers a surgical robot that places threads thinner than a human hair into tissue that moves with breathing and pulse, about three times closer to target than the scripted routine it replaced.
A quadruped learns to walk by practicing in thousands of simulated worlds at the same time.
Simulated rovers share data over a Reticulum mesh network with heavy packet loss.
A simulated fly learns to fly through the wiring map of a real fly’s brain.
A reinforcement learning and game dev harness: coding agents explore a world model in simulation and run reinforcement learning in the loop.
A small language model plays Doom in real time, more than twice as fast as the game needs.
A quadruped with leg-mounted floats walks into water, floats and walks out. Built in 40 minutes from a design posted on LinkedIn.
An AI agent read a research paper, wrote the code, and trained a drone controller that flies more accurately than the hand-written one.
A drone learns to predict the wind and keeps its package on target in gusts it never trained on.
A drone simulation split into separate programs that talk to each other, the way software runs on a real robot.
Text and image prompts drive an agent that generates 3D characters procedurally.
Founder, CEO and CTO. Building Veoveo, a self-hosted autonomy harness for physical AI, where humans, robots and agents learn, play and evolve together in simulated and real worlds.
AlienWars Gym, Maquina and Axionomy, plus the experiments in the lab above.
Principal AI and Simulation Engineer (contract). AI and simulation infrastructure, harness, training, synthetics and building world models where humans, robots and AI agents learn, play and evolve together. Developed the simulator shown in the Hivemind fixed-wing demo.
Founder & CTO. Building scalable simulation infrastructure solutions and immersive experiences centered on human-robot-AI interaction. Designing and implementing multi-agent frameworks and behavior systems.
Lead Simulation and AI Engineer. Leading development of an open source physical AI platform for sim-to-real transfer and digital twin applications, and real-time synthetic data pipelines for training embodied AI models.
Consulting: sensor characterization and visualization tools, Unreal Engine in development operations, and autonomous swarm simulation architecture for DARPA and DoD programs.
Real-time deep learning inference in a VR music MMO using TensorFlow, GStreamer and Unreal Engine 4; physics-based VR gameplay and full-body VR networking.
Designed and built a robot-crafting sandbox game with physics-simulated robots (Bullet Physics).
Gameplay, physics and networking on console titles, including Star Wars: The Force Unleashed. Simulation tools for real-time effects and previsualization on several films.
And
yet
it moves!
Proposes modeling digital twins with three primitives: objects, operations and machines. The idea became the Maquina project.
Argues for measuring how far a virtual world is from equilibrium instead of designing for perfect balance.
Describes Situation Design: areas of activity with a radius of influence, role slots and growth rules, placed above agents and encounters.
Proposes designing worlds from meanings such as “resting place” and storing memory in places and objects as well as in NPCs.
Where a small encoder model fits in games: classifying player behavior, game state and procedural content.
Actor model, behavior trees, memory systems and language models for agents in virtual worlds.
I started in console games, writing gameplay, physics and networking code at LucasArts and other studios across the industry. Since 2016 I have worked on simulation for robotics, autonomy and defense, and on the perception and learning systems trained in those simulations.
I prefer small sets of well-defined parts: typed state, explicit rules, and runs that can be replayed. I write Rust by choice, C and CUDA where performance requires it, and Python for training and tooling.
“With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.”
| Rust | Veoveo, Axionomy |
|---|---|
| C, CUDA | AlienWars Gym |
| Lean 4 | Maquina |
| PPO, PufferLib | AlienWars Gym |
| MuJoCo | lab experiments |
| Isaac Sim, PX4 | Veoveo |
| Unreal Engine | Simbotic, Shield AI |
| Unity | games archive |
| GStreamer, LibTorch | Simbotic |
| WebAssembly | AlienWars Gym, Axionomy |
| Model Context Protocol | Veoveo, Axionomy |
| Rerun | Veoveo |
Alex Rozgo
