InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization

Researchers have developed InternVLA-A1.5, a unified model for robot manipulation that integrates pretrained vision-language models with future prediction in latent space. This model enables efficient robot manipulation with preserved semantics and long-horizon execution. It achieved the best overall results on all six simulation benchmarks and showed strong compositional generalization on held-out instruction bindings.

1sec.ai