Qianxun AI study: legacy data shows 'emergent transfer' after robot hardware upgrades, with a critical threshold

机器之心 · wechat · 2026-08-04

Qianxun AI's team led by Gao Yang reveals 'emergent transfer' in cross-configuration robot learning: after hardware upgrades, legacy data becomes useful only when the new hardware's capability crosses a threshold. Experiments show that increasing new data from 4.3 to 15.6 hours, adding legacy data boosts success from 10% to 86.7%. They propose a transfer threshold and explain via gradient alignment and residual policy uncertainty. A phase-aware data collection strategy reduces new data needs from 8 to 1.5 hours in a watering task, raising success from 40% to 78.3%.

Original post →

More from Embodied

Embodied channel →