TREE NEWS update: Mifeng Technology released a study on real-machine data quality for embodied AI, drawing on over 200,000 hours of real-robot data and more than 20,000 task records. Models retained about 90% of extra detour paths from demonstrations, and task time rose from 25.3 seconds to 32.6 seconds when detour data went from 0% to 80%. In another experiment, improving visual information completeness raised correct arm-selection accuracy from about 50% to 100%.
Mifeng Study Finds Robot Models Retain ~90% of Detour Paths From Demonstrations
The finding that models absorb roughly 90% of demonstration detours is a quiet warning about data provenance in embodied AI: real-robot datasets carry operator habits and inefficiencies that models inherit rather than filter out, with measurable task-time costs. This matters most for teams scaling imitation learning on large real-world corpora, where volume alone is treated as an asset. The visual-completeness result points the other way, suggesting input quality can move outcomes more sharply than data quantity. Whether the industry's push toward ever-larger real-machine datasets accounts for this contamination is the open question.
Generated by AI for reference only.
Share on WeChat
Open WeChat → Scan → then tap "…" to send to a chat or Moments.
Tap "…" in the top-right corner to send to a chat or share to Moments.