🤖 InternVLA-A1.5 · VLA Action Explorer

Explore InternVLA-A1.5-base, a Vision-Language-Action robot policy (Qwen3.5-2B backbone + flow-matching action expert) from the paper "InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization."

Give the model a camera frame and a task instruction; it predicts the next 50-step continuous action chunk — the robot's planned motion — which we plot per dimension. This is the model's flow-matching action output (normalized units) for the google_robot single-arm embodiment; the video-foresight branch is discarded at inference.