SEED and the Art of Making Research Behave

A football stadium is an excellent place to hide a research project. The crowd watches the field. The camera follows the ball. Nobody pauses a touchdown to congratulate the indirect illumination. Yet inside EA SPORTS College Football 25, the light bouncing around more than 150 stadiums carries a small story about how an experiment becomes useful.

The team that changed its mind about light

The College Football team initially did not expect to use GIBS, short for Global Illumination Based on Surfels. Lighting does not change much during a single match. A dynamic lighting system sounded like an answer to a question the team was barely asking.

Then came a better reason. With GIBS, lighting artists and environment artists could work in tandem. The technology removed dependencies in the production process. A stadium could be built while its lighting took shape. The benefit was partly organizational: fewer people waiting for other people.

“Lighters just light and environment artists just build structures.”Jay Goodman · Technical Art Director, College Football

That observation captures SEED rather well. A beautiful image can persuade a conference audience. A better working day can persuade a game team. GIBS still needed optimization, a system to prepare its lighting during loading, and close collaboration to meet a 60-frames-per-second target on supported consoles. Adoption was earned in those details.

A laboratory with a delivery address

SEED means Search for Extraordinary Experiences Division, a name with the faint air of a government bureau commissioned to investigate delight. EA established it in 2015. The assignment is applied research: explore what interactive entertainment might become, then build enough of it to discover whether the idea holds up.

Its position inside EA matters. The immediate users are artists, engineers, designers, and game teams. Players encounter the results in a character’s expression or a goalkeeper’s movement, often without encountering SEED’s name. The research division sits between academic inquiry, engine development, and the inconvenient business of finishing a game.

There is a commercial logic to this arrangement. EA gets technology it can apply within its games and development processes. Selected tools become public. SEED’s business is internal R&D rather than a catalogue of subscriptions. Comparing it with a software startup misses the advantage of having game teams close enough to ask awkward questions.

Consider PICA PICA. The 2018 experiment combined self-learning agents, a procedurally assembled world, and real-time ray tracing in SEED’s Halcyon research engine. It demonstrated reflections, soft shadows, and other lighting effects using Microsoft’s DirectX Raytracing API. The robots were entertaining; the rendering techniques were the point.

Microsoft and NVIDIA were collaborators in that early work. The route toward production continued with Frostbite, EA’s engine team. This is a useful distinction for anyone studying the lab: demonstrating a technique and making it belong in an engine are separate jobs, with separate opportunities for trouble.

The goalkeeper has homework

In FC 26, SEED worked with EA SPORTS and Frostbite on reinforcement-learning goalkeeper positioning. The goalkeeper learns through in-game situations, while a feedback and fine-tuning framework gives designers a way to evaluate the behavior. Frostbite supplies training integration and runs the model during gameplay.

EA’s March 2026 account says training can happen overnight. These are reported comparisons for a particular system, rather than a promise that every game AI project will behave similarly. The useful feature is the loop: train, examine, adjust. A goalkeeper with a higher save rate still needs to produce football someone wants to play.

For players, the result is positioning that responds more adaptively to the action. For developers, the example suggests a narrower ambition than a universal game-playing intelligence: choose a behavior, make it trainable, and give the people responsible for the experience a way to judge it.

What the helicopter taught the researchers

A 2023 paper about testing Battlefield 2042 and Dead Space makes the engineering constraints unusually visible. The researchers added reinforcement learning to existing scripted bots for difficult navigation tasks. They could run 250 agents across five Battlefield servers, but only seven agents in Dead Space. Multiplayer architecture offered a training advantage.

The team evaluated both PPO and SAC learning algorithms. SAC reused past experience more efficiently, but unstable results made PPO preferable. They also used a small set of game-state observations instead of rendering an image for every agent. The practical choice followed the environment’s limits.

The lesson is portable, with conditions. Training needs access to a suitable environment and enough interactions. Stripping visual detail from a test range is inappropriate when the agent relies on images. A research method becomes a production method only after someone asks what information, compute, and control the actual game allows.

A jersey, a tongue, and a smaller face

SEED’s animation work is equally particular. Swish used a neural network to predict cloth shape from a character’s skeleton pose in Madden NFL 21. Its target was tight clothing, where traditional real-time cloth simulation often produces poor results. Mesh deformation and normal maps supplied the jersey detail.

It is a pleasingly modest example of machine learning. The player does not need to know a network is involved. A wrinkle needs to look right, at the moment it is visible, within the performance budget. The technology earns its place by meeting those requirements.

Voice2Face tackles recorded speech. Its 2022 research generates facial and tongue animations and maps them to rig controls, with speech-style control among its contributions. A user study reported a preference for animations that included the tongue. Even a virtual person benefits from remembering all the equipment needed to pronounce a sentence.

By 2025, SEED researchers were addressing another obstacle: high-quality speech-driven facial models can be too large for on-device, real-time use. Their knowledge-distillation work trained smaller models from larger teachers. Experiments reported memory footprints down to 3.4 MB and future audio context down to 81 milliseconds.

The compromise did not disappear. The real-time variant showed jitter. Averaging predictions across three frames smoothed it, with added latency and more computation. This is the kind of detail that makes research useful to another practitioner: the improvement arrives with a bill, and the bill names what must be paid.

Borrow the tool; keep your judgment

Outside EA, the most direct way to use SEED’s work is through its public software. Gigi lets a graphics developer describe a rendering technique as a node graph, inspect it in a viewer, and generate code. The developer still supplies shader code; Gigi handles much of the mechanical integration work.

The important design choice is code generation. Different teams have different engines and conventions. Rather than ask all of them to adopt one runtime abstraction, Gigi produces code they can inspect and integrate. Its viewer supports debugging, profiling, hot reloading, and Python automation for repetitive experiments.

Other public projects include Dem Bones for extracting skeletal skinning from example meshes and a WebGPU implementation of Position-Based MPM. These releases give students, researchers, and developers something concrete to examine. They also show a culture of sharing selected machinery, rather than merely displaying finished pictures.

That culture includes questioning the measuring instruments. SEED’s participation in the GENEA gesture-generation challenge concerns how speech-driven gestures should be evaluated. Human-like movement is an awkward subject for a convenient score. Comparing automated metrics with user judgments helps expose where apparent progress and perceived quality drift apart.

For another company, the useful habits to borrow are fairly ordinary: tackle a specific production difficulty, involve the people who will use the result, publish enough detail for a peer to inspect it, and measure what matters to the experience. None guarantees success. Together, they make an experiment easier to challenge before it becomes expensive to defend.

SEED’s distinctive position is the combination of a research remit and proximity to shipped games. A football jersey, a speaking face, and stadium lighting demand different expertise. They share one unglamorous test: can another team put this to work? In that question, the lab finds plenty to investigate.