Skild AI S1 learns tasks from videos

S1’s key shift is moving robot teaching from expensive retraining into inference-time behavior prompting—a potentially major unlock for scalable physical AI.

• –Single-video task prompting could reduce the need for robot-specific demonstrations and fine-tuning
• –The hardest problem remains the embodiment gap: mapping human motion, forces, and intent onto different robot bodies
• –The approach builds on Skild’s use of internet video, simulation, and cross-hardware training to improve generalization
• –Developers will need evidence on unseen-task success rates, latency, safety, failure recovery, and supported robot platforms