Lecture 20 and 21 - Robots Flashcards
Examples of robots
Roombas move very randomly
Infrared or bumper – when bumps into something, then turns 30 degrees
Eventually clean out whole house… thanks to very long battery life
If bumper or infrared detector blocked (like a cat’s tail on it), will just keep spinning
Da Vinci robot
Medical robot
Very expensive
Very precise…. Key whole surgeries
Low autonomy, often remote controlled by surgeon
Telepresence robot
Exoskeletons
Industrial robots
Entertainment and consumer robots
Art projects
Disaster recovery robots
What is a robot?
A robot is a machine - especially
one programmable by a computer -
capable of carrying out a complex
series of actions automatically
– Oxford Dict
Prof not fan of Oxford dictionary definition of a robot
- sensors
- actuators/effectors
- program/controller
(- embodied in environment)
No requirement on autonomy according to prof
Thinks google’s definition of robot vs. machine is bullshit
Robots don’t have to be autonomous!
Machine vs. robot:
Slighting scale of autonomy and purpose
Robot qualities
● Sensing + Acting - perceive/do?
-Sensing… infrared, fullness detector
-Acting… actuators (wheels, vacuum motor, notification of battery sensor)
● Embodied AI - cognitive functions?
-Cognitive functions… not a lot. Randomly explore until bump into something then turn right
● Environment + Goa
-Environment… flat indoor environment
-Goal… clean floor
More about sensing + acting
Sensors = input from the environment
Effector/actuators = carry out actions
Sensors:
● Perception: animals
○ eyes,
○ ears,
○ nose,
○ vestibular system,
○ skin
○ magnetics, (some)
○ echolocation (some)
● Perception: robots
○ cameras,
○ mics,
○ electronic nose,
○ accelerometer,
○ artificial skin,
○ electronic compass,
○ lidar/radar (ex: in self-driving cars)
○ GPS
○ particulate sensors
○ (absolute) temperature/humidity sensors
Actuators:
Animals:
● Communication: face, gestures
● Movement: muscles, legs
● Energy: digestive system
● Structure: bone, cartilage
Robots:
● Communication: displays, gestures, lights
● Movement: motors, wheels, wings, propeller
● Energy: batteries, photovoltaics
● Structure: metal, plastic, carbon fibre, ..
Intelligent Agent Paradigm
“Intelligence” → agent
(Russel & Norvig, 2009)
● Rationality
● Sufficiency (autonomy)
● Sustainability (persistence)
● Communicability
● Cooperation
● Mobility
● Adaptability
Intelligent agent criteria similar to David Vernon’s
Intelligent Agent Environments
*Observability… parts of the environment observable to the robot (ex: vision of self-driving car)
*Dynamics…. Enviro. Repeatable, learn model of the world, statistics things reoccurring (ex: mario bros)
*Experience… anything can learn before hand, repeatability… (ex: chess can learn rules before hand)
*Continuity… discrete (ex: left, right, jump in Mario) vs. large continuous (bunch of things you can do… like in GTA)
*Multiagent… compete or collaborate with other agent
Robots need to estimate
their current state!!!
-understand where it is in space
-proprioception
Control Loops
Cybernetic system = system that has a feedback loop
environment, sensors, controller, action… loop
-also called feedback loop or closed-loop control system
The human body is cybernetic system because self-preserving system, keeps itself alive, self-regulates
Agent puts out actions into the environment, which returns it states and rewards
Open-loop:
Environment, sensors, controller, action. STOPS. Doesn’t loop
Aka feedforward or open-loop control system
Ex: toaster and traffic light, electric hand dryer… just have a timer and then stop. Don’t check if there’s traffic or the toast is toasted or the hair is dry. They just stop
Summary so far
● Robots for all purposes: medical, telepres, exosk., manuf., art
● Robot: sensors + actuators/effectors + controller + embodiment
● Robot qualities:
○ Perception/action?
○ Cognitive functions?
○ Env. + Goal
● Intelligent Agent Paradigm + Intelligent Agent Environment → estimate own state
● Control:
○ Cybernetic aka. Feedback aka. Closed-loop control
○ Feed-forward aka. open-loop
Evolutionary
Algorithms
(aka how to make
a robot)
- Create random population of agents
- Evaluate each agent
- Pick the fittest
- Add random mutations → create N mutated copies
- Go to (2)
Evolutionary algorithm:
General framework for optimizing something
The numbers (like 0.5, 0.8) are fitness scores
Ex: how close is each candidate to the goal of a circle. fitness score ex:0.8, closest to a circle.
Examples:
-Generation 1 vs. Generation 999 of a walking robot
-In a video game too
-solid but weird shaped pieces of furniture
EA example: nasa antenna
Make a bunch of random shapes, test them, take the least bad one , mutate it..; repeat for generations
EA synonyms:
Evolutionary Programming
Genetic Algorithm (bit different)= take two best solutions (instead of 1) and create a crossover
Evolutionary Strategies
Dumb but interesting robots
Simone Giertz
Made shitty robots until became really good
Braitenberg Vehicles, 1984
18
-Positive: Light source stronger = motor goes faster = wheel turns faster
-Move the light more on one side, turns faster on one wheel, turns away from the lights “fear”
-If cross over wires, will turn towards the lights “aggression”
-Negative: Can hook them up also with a negator (if stronger light, motor has weaker velocity (decelerate), so opposite side will turn towards the light “affection” (approach gently without touching it)
● Possibly simplest example of cybernetic system
● Minimal “cognition”, but interesting behaviour
● Other simple feedback loops in animals and humans?
iCub, 2009
● Dimensions of 3.5yr old
● Actuators:
○ 53 degrees-of-freedom
-Degrees of freedom = motors (or things that can move along an axis)
○ LEDs for eyebrows+mouth
● Sensors:
○ Stereo cameras (eyes)
○ Haptic sensors (skin + fingertips)
○ Microphones (ears)
○ Joint Encoders
○ IMU
But why?
● Computationalism
○ → Embodied Cognition
-Counter movement to computationalism is embodied cognition
-Icup is an attempt at embodied cognition
● Top-down logic
○ → (Alan Turing:) Baby Machines/Child Machine = instead of programming it with everything it should know, raise the robot with humans and teach it everything you would teach a human child
-could teach it human values
-Icup trying to learn hand and eye coordination in the video
-curiosity module
Development psychologists and roboticists
iCub Successes
● Intrinsically-motivated learning
● Object recognition through curiosity
● Better tactile finger sensors
● Robot acceptance/trust
● New silicone sensor skin
● Adversarial vision attack robustness
● Symbol Grounding
● iCub cognitive architecture
Open problems
- Foundation Models for Robots
Missing:
● Large high-q annotated datasets
● Easy adaptation pipeline
● Cheap training + human biases
LLM, ChatGpt4 is a foundation model
Teach it all the words in the English language and it becomes good at a bunch of stuff
- Useful home assistants
Missing:
● Cheap+powerful robots
-each is like 400,000$
● Human understanding
● Dexterity
● Good 3D mapping
● Affordances
● Adjustment on the fly
Yelling at your robot somehow helps?
Summary of second lecture
● Evolutionary Algorithms
○ Used for robot training, robot design, material design, general optimization
○ Make random population, evaluate fitness, mutate best → create offspring, repeat
● Braitenberg Vehicles
○ Simple feedback system but interesting behaviour
● iCub
○ 3.5yr human model to study embodied cognition
○ Used for self-guided learning (intrinsic motivation), symbol grounding, and hardware dev
● Open problems:
○ Foundation models
○ Useful home assistants
