Lecture 1: Neurorobotics

La robotique peut aider les neurosciences :

• certaines des questions qu’on se pose en robotique sont très similaires à des questions des neurosciences

(switch to English, there are international students)

Ex: computational neuroscience ⟶ gives models that are then tackled from a robotics standpoint

Sometimes, tuning parameters in robotics gives rise to new models, that can then be tested by neuroscientists.

There are often issues you didn’t even think of before implementing your neuroscience model in a robot.

There several difficulties:

• perception
• decision
• memory
• action planning
• motion plannings
• etc…

then, you have to integrate all these things together (synchronization).

But in any case ⟶ Emphasis on the importance of the body

⟹ Thinking is tighly tied to the body

Usually, in computational neuroscience, experiments are very specific, and not general enough ⟶ robotics can urge us to try to generalize, since robots are confronted to unexpected environments

• Symbolic Artifical Intelligence:

  • remains theoretical, at the level of reasoning only
    • ex: Alphago is very powerful, but you’ll not encounter any robot being able to actually play physical chess/go.

• Industrial Robots

• Computational Neuroscience ⟶ very specific problems

“What I cannot creat, I do not understand” – Richard Feynman

Big Data learning (alphago, Google DeepMind’s Atari games, etc…)

VS

Micro Data Learning (how humans and animals learn: with a dozens of trials only)

Simultaneous Localization And Mapping:

use external observations to refine your estimation of your position

Bayesian filters:

• Kalman filters
• Particle filters

(cf. course by Pierre Bessière)

History of the field

• 40-70’s: First: Top-Down approach (thoeretic/high-level ⟶ applied)

• 90’s: Bottom-Up approach

Logic theorists ⟶ wanted to prove simple theorem by imitating human reasoning

AI:

Resolving problems usually solved by high-level mental human processes

Grey Walter’s tortoises:

• attracted by light sources
• stopped when bumped into an obstacle

1966-72: Stanford’s Shakey robot

John Searl’s Chinese room experiment: philosophical criticism of the Turing Test

Gibson’s affordance theory (1977)

O’Regan’s sensorimotor theory

The Animat Approach

1984: Braitenberg’s vehicles

Kenji Doya (2000):

• Cerebral cortex ⟶ Unsupervized/latent learning

• Basal ganglia ⟶ Reinforcement learning

• Cerebellum ⟶ Supervized learning

Tani \& Nolfi (1999): emergence of sensi-motor representations

Hippocampal place cells

Extrinsic motivation VS Intrisic motivation