// blog · analysis · research-papers2026-08-06source: arXiv and code releases

Twenty-one percent, and the whole-body problem

Learning robot actions from video that has no action labels in it. If the latent space transfers across bodies, it is the answer to the field's central bottleneck. If it does not, it is a very good result on one platform.

An ICLR 2026 paper presents a unified latent vision-language-action framework for whole-body loco-manipulation that learns latent actions from action-free egocentric video, reporting a 21.3% improvement over GR00T on AgiBot X2.

Action-free is the load-bearing phrase

Robot learning is constrained by paired data: video plus the exact joint commands that produced it. That requires teleoperation or instrumented demonstration, which is expensive, slow, and does not scale with anything except headcount. Egocentric video without action labels is comparatively abundant.

Learning a latent action space from unlabelled video and grounding it in a real robot is not a new idea, and it keeps returning because the payoff is enormous if it works. The 21.3% margin is real and narrow. Whether the latent space transfers across embodiments is the question that decides whether this is a technique or a milestone.

Whole-body is exactly where the gap sits

Manipulation and locomotion have each been solved reasonably well in isolation. What falls between them is everything requiring both at once: bracing against a surface to apply force, using body weight for leverage, catching balance while carrying something. Those are not edge cases. They are most of what physical work consists of.

The other answer, arriving simultaneously

Ant Group's LingBot-VLA is trained on twenty thousand hours of real dual-arm data across nine hardware configurations. Where the academic work attacks the bottleneck with architecture, this attacks it with capital.

Nine configurations is the more technically interesting number than the twenty thousand hours. Training across multiple embodiments is what separates a policy from a foundation model, and it is precisely where most cross-embodiment efforts break, because the mapping between different kinematics is not learned for free.

The field is arguing with itself productively

Buy the data or learn without it. Both approaches shipped results in the same quarter, they are not mutually exclusive, and the winner is probably whoever combines them first — a latent space learned from abundant unlabelled video, grounded with a smaller quantity of expensive paired data.

Meanwhile the infrastructure question has started appearing in the literature at all, which is its own signal. Serving systems show up when a technology stops being demonstrated and starts being depended on.

OpenDriveLab — WholeBodyVLA: towards unified latent VLA for whole-body loco-manipulation control → · arXiv — GR00T N1: an open foundation model for generalist humanoid robots → · arXiv — SmolVLA: a vision-language-action model for affordable and efficient robotics →