Researchers have built a single world model that handles robot locomotion, camera movement, and hand-joint control without fragmenting into separate systems.
Worldscape-MoE uses a Mixture-of-Experts architecture layered on top of Diffusion Transformers. The key insight is that different control signals - robot limb commands, camera trajectories, hand-joint angles - all constrain the same underlying physics, even if their representations look nothing alike. Instead of training separate models for each interface, Worldscape-MoE injects each control type through modality-aware adapters while routing computation through both shared and control-specific expert networks. The team also built in a progressive tuning strategy, so new control types can be added without retraining from scratch.
The results challenge a reasonable assumption: that mixing incompatible control signals would cause interference and degrade each individual task. Instead, heterogeneous training improved results on WorldArena and across locomotion, manipulation, and egocentric hand-control benchmarks. If that pattern holds at scale - and the paper reports early scaling signals as more data and experts are added - the field may have been paying a real cost by keeping these systems siloed.
Worldscape-MoE is a research paper, not a shipping product, so the gap between benchmark wins and a robot that actually works in a warehouse remains as wide as ever.