Robotics Field Still Lacks Key Advancements: DeepMind's Gopalakrishnan

Google DeepMind's robotics research lead, Keerthana Gopalakrishnan, has given a candid assessment of the field. Despite impressive demonstrations of humanoid robots sprinting through obstacle courses, she believes that the technology is still in its 'GPT-2 era', lacking key advancements such as cross-embodiment generalization and few-shot learning.

Gopalakrishnan pointed out that locomotion has made rapid progress due to simulation-friendly physics, but this is a low-value sideshow compared to manipulation, the actual commercial frontier. However, manipulation remains bottlenecked by contact dynamics that simulators model poorly.

The Gemini Robotics 2 architecture, developed by DeepMind, splits embodied reasoning from action control into three distinct models. The reasoning model, ER2, has been made publicly available and has seen unexpected adoption in academia for benchmarks and low-level control work.

Gopalakrishnan emphasized the importance of reliability over capability, highlighting that a robot can perform a task successfully in a demo but fail repeatedly when deployed in real-world scenarios. She suggested that tasks with retry tolerance, such as pick-and-place, are closer to deployment-ready than those without.