Robotics researcher: foundation models only cut costs, performance is the real bottleneck

Robotics researcher Glen Berseth posted a series of tweets on September 16 to cool down the recent hype around "large foundation models on robots." His core point: foundation models only advance along one axis—cost—with no true breakthrough in performance. Current conclusion: for embodied intelligence to move forward, effort must go toward raising the performance ceiling, not just pushing down the cost curve.

Confirmed

• Berseth recalled history: as early as the 1960s, Marvin Minsky's lab already had robotic arms using computer vision to recognize object geometry, plan paths, and assemble parts.
• He argues today's foundation models solve the same class of problems in robotics, only cheaper and faster to deploy—essentially shifting the cost curve leftward.
• He listed directions that genuinely need solving: more object categories, robustness in wild environments, multi-robot and human-robot interaction, real-time control, and better hardware—he specifically noted the human hand remains impressive.
• Using autonomous driving as an example: decades of investment still haven't cracked it, because such tasks require real-time closed-loop feedback, not cheaper static perception plus planning.

Why it matters

• This view provides historical grounding and a sober perspective for the hot embodied-AI narrative: cost reductions are real, but if the performance bottleneck remains, robots won't leave controlled environments.
• Berseth's argument reminds the research community and investors to distinguish "progress on the cost axis" from "breakthroughs on the performance axis"—the two mean very different things for industry maturity.

2026-09-16 ~ 2026-09-16 · 10 related posts

• Minsky's 1960s robot arm solved the same class of problems as today's foundation models — GlenBerseth · 2026-09-16
• [source] Robotics researcher: AI progress is cost-only, closed-loop control remains unsolved — GlenBerseth · 2026-09-16
• Glen Berseth: foundation models made 1960s robot tasks cheap, but performance is still the bottleneck — GlenBerseth · 2026-09-16
• Berseth: Self-driving shows cost cuts aren't enough — real-time closed-loop control is the hard part — GlenBerseth · 2026-09-16
• [source] Robotics researcher: Embodied AI progress must push performance, not just cost — GlenBerseth · 2026-09-16

5 near-duplicate retellings: GlenBerseth · GlenBerseth · GlenBerseth · GlenBerseth · GlenBerseth