InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization
InternRobotics
InternVLA-A1.5 addresses semantic drift during robot manipulation training by preserving VQA and subtask-prediction objectives alongside action learning. Future prediction is reformulated as latent-querying via learnable foresight tokens supervised by a frozen video generation model, giving the policy world-model dynamics priors without pixel-level generation. Pretrained on 1.2M robot episodes and 3M multimodal samples, it achieves state-of-the-art results on six simulation benchmarks and strong compositional generalization on real robots at real-time speeds.
Why it matters
Highest-upvoted paper on HuggingFace Daily Papers for the July 6–8 window (463 upvotes); demonstrates that latent foresight tokens from frozen video models can transfer world-model priors to a policy without the inference cost of video generation, a practical advance for real-robot deployment.
Importance: 4/5
463 HF Daily upvotes (+1 for ≥100); SoTA on 6 simulation benchmarks; practical robotics advance with public code and weights.