Generative AI trains humanoid robots from synthetic videos alone

Training humanoid robots traditionally relies on expensive, time-consuming real-world data collection. A new framework from Tsai et al., accepted to IEEE/ASME AIM 2026, cuts that dependency entirely by using generative AI to produce synthetic video demonstrations. The system takes a textual description of a task—like “grasping an object from the left side”—and generates multiple realistic, diverse human body movement sequences. These serve as training data, allowing the robot to observe and learn different execution styles without human intervention.

The method was evaluated across four simulation scenarios. Results show the robot completes tasks successfully and exhibits strong adaptability to complex motion variations, even without ever seeing real-world examples. This approach could accelerate humanoid robot training by removing data collection bottlenecks and enabling rapid skill acquisition for a broad range of tasks. It also opens the door to learning motions that are difficult or unsafe to capture from human demonstration.