Robotics & World Models — Embodied AI and Perception-Action Models

Robotics experts reflected on why household robots remain rare and described recent progress. Google’s Gemini Robotics 1.5 integrates 'think-then-act' VLA methods: robots generate textual plans before motor execution and expose internal reasoning. Motion Transfer enables cross-platform skill sharing so different robot bodies can reuse data. Data scarcity is the primary bottleneck; teams scale with simulation, YouTube video, and synthetic data. Touch/force modalities gain importance for dexterous tasks; vision suffices for simple grippers but touch is essential for fine manipulation. Beijing智源’s Emu3.5 proposes a unified multimodal world model predicting 'next-world state' tokens across image, video and text.
Key insights
Think-then-act VLA improves interpretability and multi-step task performance
Google DeepMind’s robotics group described moving from direct perception-to-act policies toward Vision-Language-Action (VLA) agents that produce explicit textual plans prior to motor commands. This approach yields two benefits: improved multi-step task decomposition and human-readable internal states that aid safety and debugging.
Motion Transfer enables cross-body skill reuse
Teams reported Motion Transfer techniques that allow skill data collected on one robotic platform to inform control policies on different embodiments. This reduces per-robot data costs by enabling shared representations and transfer learning across robot types.
Data is the primary bottleneck; simulation and synthetic video are scaling paths
Robotics leaders emphasized the high marginal cost of physical robot data and the need for scalable sources: large-scale simulation, leveraging public video (e.g., YouTube) and synthetic video generation to create diverse training scenes and behaviors.
Tactile sensing becomes critical for dexterous manipulation
While vision suffices for many coarse grasp-and-place tasks, leaders reported touch/force feedback is essential for fine manipulation with dexterous hands—tasks like scissors or delicate object alignment require haptic modalities in the control loop.
Emu3.5 and world-model approaches aim to predict 'next-world state' across modalities
Beijing institute released Emu3.5, a multimodal world model that tokenizes video, image and text into unified tokens and trains to predict the next world-state token, aiming to unify perception and dynamics across modalities and improve physical reasoning.
Near-term roadmap: domain-specific rollout before household ubiquity
Leaders predicted 2–3 years of further capability gains that make generalizable robots practical in verticals (manufacturing, logistics), with home robotics arriving later. The timeline emphasizes progressive deployment by domain rather than immediate household robots.