LLMs outputting JSON are too slow to control robot joints in real time, argues researcher

nikolaj2030 · x · 2026-08-20

The author questions how big LLM-controlled robots can get: without an actual neural net controlling each joint in real time, relying on LLMs emitting JSON introduces high latency—API responses take many milliseconds, making tasks like juggling three objects hard to imagine.

Lacking real-time tactile feedback is another bottleneck: an LLM would likely struggle to grasp an egg without breaking it and crack it into a pan, perhaps only succeeding extremely slowly (20+ seconds). But for tasks needing neither fast reactions nor touch, he expects LLMs to be quite capable within a couple of years—e.g., folding laundry.

The quoted tweet adds a benchmark: Opus 5 completed a robot task with zero demonstrations (despite no robot training) but took 7 minutes, while GEN-1.5 took 7 seconds.

Related event: Skeptic argues LLM-controlled robots have under 10% chance of reliability by 2028(2 posts)→

Original post →

More from Embodied

Embodied channel →