Lumo 2 Context

Specifications

System Type: Latent world-action robotics foundation model
Training Pipeline: Three-stage progressive modality pre-alignment and joint training
Control Scope: Predictive latent-dynamics robot control
Training Data Source: Vision-language corpora; egocentric human video; multi-robot demonstrations
Supported Platforms: Cross-embodiment robot platforms and human-to-robot transfer
Evaluation Mode: Embodied benchmarks and real-world manipulation
Tasks Demonstrated: Pick-and-place, temporal reasoning, long-horizon and dexterous manipulation
Release Model: Research technical report