One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment & Humanoids

The current state of robotics research balances rapid progress in simulation-based locomotion with the persistent challenge of achieving reliable, general-purpose manipulation in the real world. While humanoid robots can now perform impressive feats like high-speed running, practical utility remains limited by the need for robust, cross-embodiment generalization and safety. Google DeepMind’s Gemini Robotics 2 suite, including the Embodied Reasoning (ER2) model, represents a shift toward integrating high-level semantic reasoning with low-level whole-body control. Despite advancements in few-shot learning and multimodal interaction, robotics currently occupies a "GPT-2 era," where models lack the universal reliability and cross-platform versatility seen in digital LLMs. Future breakthroughs depend on refining data strategies—balancing teleoperation, sensor-based UMI, and human-centric video—while addressing the critical need for safety, physical robustness, and efficient orchestration of multi-robot systems.