What robots can (& can’t) do in 2025 | Ken Goldberg (UC Berkeley, Ambi Robotics)

Robotics faces significant hurdles in perception, control, and physics, particularly regarding the unpredictable nature of physical manipulation like grasping objects. While generative AI shows promise, replicating a "ChatGPT moment" in robotics remains difficult because the necessary large-scale, grounded training data for physical tasks is scarce. Ken Goldberg, a professor of engineering at UC Berkeley and co-founder of Ambi Robotics, argues that the next decade of progress depends on a hybrid approach. This strategy combines cutting-edge generative models with "good old-fashioned engineering"—such as control loops and edge detection—to bridge the gap between simulation and reality. Furthermore, current humanoid robots remain constrained by high costs, safety risks, and limited dexterity, suggesting that specialized, task-oriented robotic systems will continue to drive practical automation in logistics and manufacturing for the foreseeable future.