What is General Intuition's new approach to training robots?

The startup proposes training physical AI foundation models using millions of hours of video game simulation data. By conducting large-scale reinforcement learning in virtual environments, robots undergo countless trial-and-error iterations in a sandbox, mastering physical interaction logic before migrating to real hardware, shifting from small-sample learning to large-scale pre-training.

What does this approach mean for the robotics industry?

It tackles two major pain points: data silos and the sim-to-real gap. Traditional methods require collecting data separately for each task, while this approach leverages physics engines built into Unreal Engine and similar tools for one-train-many-scenario reuse. Hardware manufacturers that fail to integrate AI training capabilities risk being overtaken by software-defined hardware.

What key signals should we watch for going forward?

The core signal is whether sim-to-real transfer accuracy can break through to practical levels. Investors should monitor benchmark test data, especially zero-shot or few-shot generalization in unseen physical scenarios. The depth of collaboration between game engine vendors and AI startups is also a critical variable.

這家初創公司認為，機器人領域即將迎來它的 ChatGPT 時刻

General Intuition 公司押注數百萬小時電子遊戲模擬數據可用於訓練物理AI基礎模型，透過在虛擬環境中進行大量模擬訓練，大幅降低對真實世界數據採集的依賴，從而更輕鬆地構建更加智能的機器人系統。

Sources

TechCrunch AI