World Dreamer v5-lite gave several robots one shared world state and hit 74.0 percent on RoboFactory
AspireAI community · arXiv
Most vision-language-action and world-action models drive one robot. Combining views from several robots is not enough on its own, because nothing in that combination establishes where one robot is relative to another. World Dreamer v5-lite builds a shared world state in workspace coordinates, fills it with calibrated features from a pretrained video encoder, and gives each robot its own query against that state while keeping their action streams separate.
Training withholds selected camera views from the visual branches while keeping every action target, which pushes the model toward using what other robots can see. Pretraining used more than 10,000 hours of trajectories from 20 simulated robot types, plus RoboFactory and real robots. On RoboFactory it reached 74.0 percent mean success, 23.7 points above the strongest evaluated baseline; across four real-robot tasks it reached 78.8 percent, a margin of 13.8 points, and 58.08 percent overall on RoboCasa365.
The honest limitation is that the baseline it beats is one the authors chose, and the margins are reported against a single strongest baseline rather than across a field. The world state is also fixed in spatial size as views are added, which is a design choice that will need revisiting as the number of robots grows.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.