Foundation models just shift the cost curve left — robotics should aim higher

GlenBerseth · x · 2026-09-16

Robotics researcher Glen Berseth wrote a short article framing recent progress in applying large foundation models to robotics tasks, clarifying where they succeed and where progress is still needed.

Robotics is not new: Minsky's lab demonstrated vision-guided arm manipulation and assembly decades ago. Recent foundation models do improve performance on fixed-scene manipulation while cutting developer time to produce such solutions.

But he asks which direction the field should take for flexible automation. Since high performance on fixed-environment tasks has existed for decades while the automatable space is far larger — self-driving improved slowly, largely due to the need for closed-loop feedback to adapt quickly to environmental changes. Citing Jitendra Malik — "robotics is also about dexterity and dynamics" — he argues the goal shouldn't just be moving the cost curve left, but pushing performance up: more object variety, robustness in the wild, multi-robot and human-robot interaction, real-time control, and better hardware (human hands are still impressive).

Related event: Robotics researcher: foundation models only cut costs, performance is the real bottleneck(10 posts)→

Original post →

More from Embodied

Embodied channel →