Gemini Robotics 2: 'Cross-Embodiment + Long-Horizon Tasks' and 'Layered Architecture' - 0731

English summary

In our July 18 WAIC insights (Part 1), we highlighted that 'cross-embodiment + long-horizon tasks' is a high standard for evaluating the generalization of embodied models. Recent releases, including GEN-1 (released on July 23, targeting 'cross-embodiment' for dexterous hands) and Zhiyuan τ0-VLA (released on July 27, targeting 'long-horizon tasks'), confirm that the industry is moving toward these two trends. The Gemini Robotics 2, released on July 30, further strengthens the 'cross-embodiment + long-horizon tasks' direction. In addition, the 'layered architecture' idea mentioned in our weekly report last week is also significantly reflected in Gemini Robotics 2. Based on personal understanding, I briefly elaborate on the highlights of DeepMind's three models in terms of architecture and capabilities: Model architecture: layered. Although there are 3 models, they still form a two-layer architecture: Gemini Robotics ER 2 handles high-level semantic subtasks, and Gemini Robotics 2 handles low-level continuous action generation. ...