A research team has released HiMoE-VLA, a framework designed to fix a persistent flaw in how generalist robot AI is trained.
When you train a single robot policy on data from many different robots — with different limbs, sensors, and action vocabularies — the model can learn to perform worse than if it had trained on less data. Researchers call this negative transfer. HiMoE-VLA attacks it with a layered architecture: specialized "mixture-of-experts" layers handle the input and output boundaries where action spaces diverge, a middle tier balances capacity across variations in scenes and hardware, and dense Transformer blocks in the center integrate the shared representations. Two auxiliary training objectives keep the hierarchy honest — one sharpens boundary specialization, one prevents any single expert from hogging compute.
The results are specific enough to take seriously. The model scores 3.98 on the CALVIN benchmark, 98.0% on LIBERO, and posts 75.0% and 63.7% average success rates on real physical tasks with xArm7 and ALOHA robot platforms. More pointedly, the paper claims HiMoE-VLA converts the negative transfer seen in strong baselines into positive transfer under controlled heterogeneous co-training — meaning more diverse data actually helps rather than hurts.
Generalist robot policies have become a crowded research space, with labs racing to match or beat human-curated single-task models using cheaper mixed datasets. If the negative-transfer fix holds up across broader evaluations, the harder problem shifts from "how do we gather robot data" to "how do we keep mixing it without degrading performance" — and this paper is an early, open-source answer. Code and models are on GitHub.