Gemini Robotics 2: 'Cross-Embodiment + Long-Horizon' and 'Hierarchical Architecture' - 07/31

English summary

In our WAIC observations (Part 1) on 07/18, we highlighted that 'cross-embodiment + long-horizon' are key criteria for generalization of embodied models. Recent releases, including GEN-1 (cross-embodiment for dexterous hands) on 07/23 and Agibot τ0-VLA (long-horizon) on 07/27, confirm the industry is moving toward these two trends. Gemini Robotics 2, released on 07/30, further strengthens both directions. Additionally, the 'hierarchical architecture' concept we mentioned in last week's report is also evident in Gemini Robotics 2. Based on our understanding, here are the highlights of DeepMind's three models in terms of architecture and capabilities: Architecture: hierarchical. Although there are 3 models, it is essentially a two-layer architecture: Gemini Robotics ER 2 handles high-level semantic subtasks, while Gemini Robotics 2 handles low-level continuous action generation.