GLIDE has an LLM write failure-filtering guardrails, lifting robot task success from 0% to 70%

stepjamUK · x · 2026-09-23

Some robot tasks are impossible — not just hard — for humans to teleoperate cleanly, like carrying a plate across a table with two independently controlled arms. GLIDE takes a novel approach: an LLM reads the task, predicts where a human demonstrator will fail, and writes a filter between commands and the robot to prevent it. It then watches the data, catches remaining failures, and rewrites itself.

Results: the impossible task jumps from 0% to 70% success; two others go from 10 and 0 up to 90 and 90. Keeping the guardrail at deployment lifts a barely-working policy from 0 to 60-70%. The takeaway: much of "the robot can't learn this" is really "no human could demonstrate it in the first place."

Original post →

More from Embodied

Embodied channel →