General Intuition Raises $220M to Train Robots on Billions of Gaming Videos

General Intuition has raised another $220 million at a $6.2 billion post-money valuation as it begins offering its AI models outside the company. The New York startup trains systems to choose control inputs in games, software and robots, an approach it calls action foundation models.

The pitch matters because robotics labs depend on costly demonstrations collected from human operators. General Intuition says gameplay from Medal, the clip-sharing platform from which it spun out, provides billions of frames paired with the buttons and mouse movements that produced them. That dataset could offer a cheaper route to training general-purpose agents, although the company has yet to publish independent benchmarks or detailed access terms.

Medal turns gameplay into labels

General Intuition spun out of Medal, which is also led by co-founder and CEO Pim de Witte. Medal supplies hundreds of millions of hours of gameplay in which video frames can be aligned with recorded keyboard, mouse or controller inputs.

Ordinary internet video shows what happened but rarely includes the exact action that caused each movement. Medal's input records provide those labels, allowing a model to learn which action a person would take after seeing a particular scene.

Medal is on track to receive three billion uploaded videos annually, according to the company. General Intuition says leading world-model and robotics labs train on less than 1% of the action data available to it, though it has not disclosed the dataset's composition, filtering process or share of clips with usable input labels.

Valuation jumps to $6.2 billion

General Intuition has now disclosed $674 million in funding across three rounds. Its valuation rose from $2.3 billion after the Series A to $6.2 billion after the latest investment.

Valor Equity Partners, Point72 Ventures and Seven Seven Six joined as new investors. Khosla Ventures and General Catalyst participated again, following earlier backing from Jeff Bezos through Bezos Expeditions and former Google CEO Eric Schmidt through Hillspire.

Predicting the next move

An action foundation model receives visual observations and predicts the control input a person would make next. Depending on the environment, that output could be a key press, mouse movement, controller command or robot motor instruction.

The term “foundation model” signals that the system is pretrained across many environments and can then be adapted to new tasks. General Intuition claims its models can operate in real time inside environments absent from their training data, but the company announcement does not provide standardized results for that claim.

The company also trains world models that predict future video frames from the current scene and a proposed action. An action policy can use those generated futures to practice decisions inside a simulated environment before controlling software or hardware.

Games provide the training ladder

General Intuition trains first in games where every visual observation can be paired with the player's input. It then transfers the learned behavior into increasingly realistic simulations and real-world footage, where ordinary videos lack equivalent control records.

This curriculum aims to close the “reality gap” between games and physical machines. Success depends on whether the model learns reusable concepts such as motion, navigation and cause and effect, rather than game-specific visual patterns or control conventions.

Eight minutes, one quadruped

General Intuition reports that one model played a Fortnite-style game continuously for 100 hours. It also says the model controlled a quadrupedal robot through a single front-facing camera after fine-tuning on eight minutes of real-world data.

Those demonstrations have not been independently replicated, and the company has not released enough methodology to compare them with established vision-language-action systems. Open questions include how much simulation data preceded the eight-minute fine-tuning run, whether the robot encountered unseen terrain and how frequently a human had to intervene.

Gameplay meets the robotics bottleneck

General Intuition is competing with Physical Intelligence's pi-series models, Google DeepMind's SIMA agents and NVIDIA's GR00T platform. Each project seeks reusable models that can perceive an environment, interpret a goal and produce actions across multiple tasks.

Gameplay gives General Intuition a distinctive source of action-labeled data. Robotics demonstrations require hardware, operators and controlled collection sessions, while Medal continuously receives recordings from people already interacting with complex virtual environments.

The domain gap remains the central technical risk. Games simplify physics, expose a limited set of controls and often use camera angles or visual cues that do not resemble a robot's sensors. Strong game performance therefore does not establish reliable control of physical systems.

Access remains the missing spec

General Intuition says its models are beginning to become available externally, but it has not published an API, SDK, pricing, license or public documentation. Developers also lack basic deployment details, including supported inputs, action schemas, latency, hardware requirements and options for fine-tuning.

The company plans to use the new capital for compute infrastructure, research and engineering hires, and physical AI development. A strategic partnership with CoreWeave will support its computing requirements.

Developers evaluating the platform will need answers to several concrete questions:

• Will access come through hosted inference, downloadable weights or private deployments?
• Which games, software interfaces and robot embodiments are supported?
• How are continuous robot controls represented alongside discrete keyboard and controller actions?
• What data and compute are required to adapt the model to a new task?
• How does performance compare with vision-language-action baselines on independent benchmarks?
• What safeguards handle uncertain actions, distribution shifts and physical failures?
• How were Medal clips licensed, filtered and checked for privacy or sensitive content?

The tests that matter next

Independent evaluations will need to measure transfer into unfamiliar visual settings, recovery from failed actions and performance on manipulation tasks beyond quadruped locomotion. Researchers will also need ablations showing how results change with less gameplay data, fewer action labels and different amounts of real-world fine-tuning.

General Intuition's funding gives it the resources to test whether gameplay can supply reusable training data for embodied agents. Public documentation, reproducible benchmarks and working developer access will determine how much of that approach can move from controlled demonstrations into production systems.