Training vision-language-action (VLA) policies for humanoid loco-manipulation usually depends on large amounts of human teleoperation data, which is costly and complex to collect. A new arXiv paper introduces LEGS, a method designed to fine-tune such policies without teleoperation, using an embodied Gaussian splatting world as the training environment.

The abstract notes that simulator fine-tuning of VLA policies has so far been limited, though the sentence is cut off in the available source. The title indicates that LEGS leverages a Gaussian splatting representation to create an embodied world for training. The paper argues that this approach can reduce the burden of demonstration collection.

Since only the abstract's first sentence is fully available, this article cannot assess experimental results or comparisons. The source is a single paper, so no independent verification or contrasting views are presented.