Skild AI unveils S1, a robotics foundation model that it says can learn tasks never seen during pretraining, using a single video demo, without fine-tuning (Skild AI)
UNSEEN TASKS10-MINUTE HORIZONSONE VIDEO PROMPTNO POST-TRAINING — The evolution of language modeling provides a blueprint …
Skild AI has introduced S1, a robotics foundation model capable of learning tasks not seen during pretraining using just a single video demonstration, without any fine-tuning. This development follows the evolution of language modeling, which saw the shift from BERT to GPT-3, where models gained the ability to learn from just one or a few examples.
Skild AI's S1 is designed as an in-context learner for robotics, enabling it to execute tasks after observing a video demonstration. The model's pre-training across diverse tasks enables it to understand the intent behind the demonstration, allowing it to perform unseen tasks without further fine-tuning. In-context learning for robotics aims to perform a task by demonstrating it, similar to how humans learn tasks through demonstration rather than language alone.
The hard part is evaluating a model's in-context capabilities, as it depends on the context of the task in relation to the training data and the duration of the task. Skild AI's S1 has demonstrated in-context learning on extremely long-horizon tasks that were never seen during pretraining, marking a significant step forward in robotics and AI.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
