Uploaded May 2020 | Updated September 2026, 1 hour ago
NAO by Maxtronics
|
4 : Notifications de NAO (French version)
🔇 TAP TO UNMUTE
![[AI Lab] Pepper Identifies obstacles](https://i.ytimg.com/vi/ReTvuCRQlq0/mqdefault.jpg "[AI Lab] Pepper Identifies obstacles
This video realized by the AI Lab at SoftBank Robotics Europe shows how state-of-the-art Deep Learning technologies can be applied to identify traversable terrain from Peppers RGB cameras.
Three approaches are evaluated: pixel-wise segmentation, pixel-wise depth estimation, and full image classification.
The first two approaches are based on off-the-shelf Deep Nets trained on the SUNRGBD and NYUv2 databases. They produce interesting results despite no fine-tuning to Peppers hardware.
The last approach uses a Deep Net adapted from ALEXNET and trained on a database built inside SBRE offices with Peppers cameras.
Each approach is successively evaluated by manually moving Pepper. Then autonomous displacements based on the full image classification are presented.")
