Robotics & embodied world models: Gemini Robotics, Emu3.5 and robot data

Robotics integration with world models advances via multimodal training, motion transfer and simulation-driven datasets.
Key insights
Shane Legg places 50% probability of minimal AGI by 2028
Shane Legg stated a 50% probability that a minimal AGI—able to perform typical human cognitive tasks—could be achieved by 2028. He emphasizes the need for improvements in sustained learning, visual reasoning and episodic memory, and supports embodied interaction and ethics-as-reasoning approaches for mitigation.
DeepMind robotics integrates LLM planning and RL control
DeepMind's robotics research described combining language-capable planners with RL-based low-level controllers: a 'big brain' for planning (VLA/GPT-style) and a 'small brain' for motor control. Gemini Robotics 1.5 introduces text-based internal reasoning, motion transfer across bodies, and reliance on large-scale simulated and video datasets to scale robot generalization.
Emu3.5 proposes unified multimodal 'next-state prediction' world models
Emu3.5 from Beijing institutes unifies text, image and video into a token stream trained to predict the next world state (visual or textual). The architecture frames next-token prediction as forecasting the world's next second, enabling cross-modal temporal continuity for prediction and planning.
Data and simulation remain primary bottlenecks for robotics scale
Robotics researchers point to data collection cost as a major barrier; solutions include scalable sim data, transfer from diverse robot bodies (motion transfer), and synthetic video generation to bootstrap training datasets for embodied models.