InternVLA-A1.5 Reports Wins on All Six Robot Sim Benchmarks

A robot policy paper that landed on arXiv this week is worth reading for the design choice more than the leaderboard line. InternVLA-A1.5 keeps its underlying vision language model doing what VLMs are pretrained for, VQA and subtask prediction, and bolts on what the authors call a lightweight unified expert for continuous action generation. Their framing is that existing unified robot models tend to erode the semantics of the pretrained backbone the moment you push them toward physical control, and this design is trying to stop that erosion.

The more novel piece is how future prediction is handled. Instead of asking the model to generate future video frames, which is expensive and pushes the network toward pixel accuracy rather than task-relevant dynamics, the authors recast prediction as a latent querying problem. A small set of learnable foresight tokens condenses the task-relevant future into a compact latent code, supervised by a frozen pretrained video generation model. At inference the video branch is discarded entirely, keeping real-time control. If you have been watching the world-model-for-robotics thread, this is a plausible way to inherit dynamics priors from those pretrained video generators without paying the pixel-generation cost every step.

The reported numbers are ambitious. Pretrained on 1.2 million robot episodes and 3 million multimodal samples, the authors say InternVLA-A1.5 achieves the best overall results on all six simulation benchmarks it was evaluated on, and that the preserved semantics deliver the strongest compositional generalization on held-out instruction bindings in the real world. Take those the way you take any self-reported robot policy result, as reported not settled: this is an arXiv preprint submitted July 6, 2026, not a peer-reviewed comparison, and the six benchmarks are not named in the abstract itself.

The honest caveat is that the abstract does not tell you which simulators, how large the real-world evaluation was, or how the model behaves on the failure modes that break composition — unusual object geometry, long chains of dependent subtasks, occlusion. It also does not spell out the compute budget or the mix behind the 1.2M robot episodes, which matters when the story turns on data curation as much as architecture.

What is worth watching is the direction. If foresight tokens supervised by a frozen video generator turn out to be a general recipe, teams working on manipulation policies get a way to buy world-model behavior without paying the video decoder tax at inference, and the pile of pretrained video generators built over the last two years starts to look like useful supervision rather than just a demo surface.