GEN-0: a general embodied foundation model

GEN-0 is an embodied foundation model that learns navigation and manipulation skills once, then transfers them to many robots and environments.

What makes it different

• Trained on multimodal sequences: vision, proprioception, language, and actions.
• Can output high-level plans plus low-level controls, switching based on robot capability.
• Uses simulation-to-real alignment so behaviors survive outside the lab.

Early capabilities

• Zero or few-shot pickup, placement, and simple tool use across room layouts.
• Natural-language commands mapped to action graphs like open the drawer, grab the spatula, place on counter.
• Safety layers to halt on unexpected force feedback or occlusions.

Why it matters

• Reduces per-robot training cost; one model can serve fleets.
• Speeds up deployments in logistics, household assistance, and labs.
• Creates a testbed for alignment and safety in physical agents.

Challenges remain: data quality, sim-to-real gaps, and clear guardrails for operation near humans. But GEN-0 shows how a single model can generalize useful skills across robots.