Robot uses toddler-like self exploration for the development of body representations
summary
Robot explores with Nao arm babbling, collects joint angles, target positions and electrical currents, records about 10,000 samples, and trains a feed‑forward neural network in Python using Lasagne to predict motor currents from positions. The model fits test data well, flags mismatches as surprise (touch, added weight, aging components), training happens offline on a laptop, and results show useful detection though not yet object identification or full online adaptation.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:43Machine learning in robotics
- 1:03Precocial vs altricial development
- 2:52Nature’s lessons for robotics
- 3:12Learning robots examples
- 3:54Experiment setup and motor exploration
- 5:25Left arm babbling data
- 6:31Neural network training
- 7:32Model training and results
- 8:58Detecting prediction surprises
- 9:42Summary and future
- 10:23Q&A
related talks
43mBuilding Qrio, a Bot That Plays Videos for My Toddler
18mRobotics: why now?
43mBuilding the Duckling Droid: AI, Robotics, and Computer Vision in Action
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
01h 35mDeep Robotic Learning
29mBridging AI & Robotics: Exploring the LeRobot Library
21mAI, Embedded System and Robotics
16mKeynote: Robot Foundation Models