![[AI Lab] Pepper Identifies obstacles @MaxtronicsRobotics](https://i.ytimg.com/vi/ReTvuCRQlq0/maxresdefault.jpg)
Uploaded September 2017 | Updated September 2026, 1 hour ago
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 Pepper's 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 Pepper's hardware.
The last approach uses a Deep Net adapted from ALEXNET and trained on a database built inside SBRE offices with Pepper's cameras.
Each approach is successively evaluated by manually moving Pepper. Then autonomous displacements based on the full image classification are presented.
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 Pepper's 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 Pepper's hardware.
The last approach uses a Deep Net adapted from ALEXNET and trained on a database built inside SBRE offices with Pepper's cameras.
Each approach is successively evaluated by manually moving Pepper. Then autonomous displacements based on the full image classification are presented.
![[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.")
