USC CSCI 445 Mid-Term 1

Polygon of Support

Draw a polygon from the points of a robot that touch the ground. If the center of gravity falls under the polygon, the robot will be statically stable

Brushless Motors

Electronic Commutation High-voltage, high-torque, expensive No firction / wear of brushes

F equation for gear

F = T / r r is radius

What does a Control Architecture Provide?

A set of organized principles for organizing a control system Structure, Constraints

State

A sufficient description of the system. Comes in EXTERNAL and INTERNAL Can be discrete, continuous, observable, hidden, or partially observable

What is a Robot?

A system which exists in the PHYSICAL world and AUTONOMOUSLY SENSES its environment and ACTS in it to achieve some goals

Degree of Freedom

A way in which a motion can be made

Reconstruction (with respect to state)

"Given the sensory reading I'm getting, what must the world be like to make the sensor give me this reading?"

How to overcome limitations of sonar

Add textures to objects (with wavelength similar to sonar) Use clever post-processing Use more advanced sensors (higher frequencies, phased array emitter/detectors)

Forward Kinematics vs. Inverse Kinematics

Forward: Given joint angle, calculate position of end-effector Inverse: given end-effector position, calculate joint angles

Problems with gears

Friction, slipping, backlash

Transmission

Gears, Belts, Chains, Pulleys Uses: Convert Power, Change axis of rotation, Motor not located at joint

Difference between maintenance and achievement goals

Maintenance: Drive 0.5 m/s Achievement: go to the end of the maze

Brushed Motors

Mechanical Commutation Low-voltage, low torque (the outside poles stay the same, but the inside poles change)

Do control architectures constrain expressiveness?

No They do affect and constrain the structure of the robot controller (e.g. behavior representation, granularity, time scale, etc.)

Do sensors provide state?

No, state is achieved after the info is processed. I.e. Have I bumped into something? Well, bump sensor activated --> process --> must have bumped into something

Horizontal Systems

Only the first layer of processing modules have access to sensor data Only the last layer of processing modules have access to actuators

Motor Power

P = w*T

Types of sensors based on energy emission

Passive: Receive energy (vision, hearing,etc.) Active: Emit energy (sonar, ladar, etc.)

How DC Motors Work

Permanent magnets with loops of wire inside, through which a current is applied

Stereo Vision vs. Ladar

Pros of Stereo Vision: dense 3D data range, high sampling rate Cons of Stereo Vision: Requires textured surfaces, incomplete range maps, limited range and depth-of-field

Perception

Provides information about its environment. Is different from sensing in that it is the sense made from sensory data. Consists of: 1. Sensing 2.Signal Processing 3. Object Recognition, etc.

How are DC/Servos motors controlled?

Pulse Width Modulation w = kv * Vavg So, the longer we provide a pulse, the greater Vavg is

What is one way to tell direction with an optosensor shaft encoder?

Quadrature encoding: use two sensors 90 out of phase

Problems with sonar

Rays may graze objects, with increasing angle of incidence we get decreasing strength of reflection, objects may be mirrors

Time Scale Differences between architectures

Reactive: Respond to the real-tie environment (short time scale) Deliberative: Look ahead (long time scale) Hybrid: Combine short and long time scales (three layer architecture) Behavior-Based: Bring the time scales closer by distributing slower computation over concurrent behavior models

Optosensor uses

Reflectance (Shaft encoding, object proximity, feature detection) Break-Beam (object presence, shaft encoding)

In addition to Looking Ahead and Time Scale, what is the third key feature of Architecture?

Representation which is the form in which the control system internally stores information

DC Motors and speed/torque

They are High speed, Low torque

What kinds of stuff can modeled (represented internally)?

Topology, objects, actions, self/ego, intentions (goals), symbols

Differential Steering

Two wheels can be be steered independently Still not holonomic

Different ways to Measure Distance with Sensors

Ultrasound: Give distance directly Infrared: Provides signal intensity Stereovision: Gives distance/depth Perspective Projection: 1 camera

How to ignore ambient light interference with an optosensor

Use Modulation of Source: Sample once with emitter off Sample again with emitter on Subtract to yield ambient value

Default Actions

Used by reactive robots in situations that don't trigger specific exclusive behaviors Since it would suck to hand-design all possible scenarios

Integral Control

Used on accumulated (steady state) error Acts in proportion to the accumulated error o = Ki*Integral(i*dt)

Proportional Control

Used on error Acts in proportion to the error, has an output: o = Kp*i, where K is a constant gain and i is input

Derivative Control

Used on oscillations Acts in proportion to the rate of change of the error o = Kd*di/dt

What is odometry?

Using the velocities of each wheel to estimate the robot position and orientation (see picture on slide 17, lec 6)

Shaft Encoder Placement

Velocity encoders are placed before the gearbox (on the motor shaft) - since higher speed yields a better resolution Position encoders are placed after the gearbox (on the output shaft) - since no backlash means better accuracy

Things that affect gains

Velocity profile of a motor, backlash, friction in the gears, surface, air, etc.

How does Back-EMF work in analog velocity encoders?

We spin the motor, then stop applying the current for a short while and measure the voltage - then infer the speed.

Holonomicity

When the number of controllable DOF matches the total number of DOF If the number of controllable DOF is greater than the total number of DOF, the robot is said to be REDUNDANT

Linear speed equation for gear

v = w*r

Relationship between meshing gears

v1 = v2 w2 = ( r1 / r2 ) * w1 T2 = ( r2 / r1 ) * T1 Gearing down means decreasing speed/ increasing torque

Relationship of Speed to Torque

w = kv*v T = ki*l Higher speed, means lower torque Lower speed, means higher torque

Control Theory is used for what?

Low-level maintenance goals

A robot is Capable of:

1. Acting Autonomously 2. Achieving Goals

Types of Perceptual System Design

1. Action Oriented: seek stimuli associated with particular scenarios 2. Expectation-Based: USe knowledge about the world as constraints on sensor interpretation 3. Focus-of-Attention: Look at what is important 4. Perceptual Classes: Partition the world into useful categories

Sources of Distinction for Architectures

1. Behavior Representation: How is action represented? 2. Granularity of Behavior: What time scale is used for action? 3. Behavior Interaction and Coordination: How are actions/behaviors chosen? 4. Basis for behavior specification: Is a biological model used? 5. Programming Methods: Is software reusable, supported, etc.

Types of Drives

1. Differential - rotation by speed of wheels 2.Synchronous - can steer wheels 3. Tracked - tanks 4. Car style

Two parts of a manipulator

1. Joints 2. Links

Two main types of robotics

1. Mobile 2. Manipulator

Two basic uses of effectors

1. Moving the robot (locomotion) 2. Move another object (manipulation)

4 characteristics of a robot

1. Physical 2. Autonomously 3. Senses 4. Acts

4 types of Control Architecture

1. Reactive (don't think, react) 2. Behavior-Based (think the way you act) 3. Deliberative (Think hard, act later) 4. Hybrid (think and act independently, in parallel)

Limitations of Odometry?

1. Requires intialization 2. Subject to curve errors 3. Assumes: Constant wheel diameter/separation, flat surface, no slipping

Examples of Bio-Robots

1. Robotically controlled Rat 2. Cockroach guidance system

A robot Consists of:

1. Sensors 2. Effectors (Actuators) 3. (Communication) 4. Controller

Two types of stability

1. Static (achieved through mechanical design) 2. Dynamic (achieved through control)

How much of the human cortex is devoted to vision?

30%

DOF Car examle

A car has 3 DOF: x,y, and theta Only 2 are controllable (x and theta) So there are more motions (parallel parking) than there are controls

State Space

All possible states the system can be in. An important point is that sensors do not provide state

Sensor Space

All possible values of sensory readings

Vertical Systems

All processing modules have access to sensor data Processing modules have access to actuators

Open-Loop (feed forward) Control

Assumes that given some applied voltage, the robot will always go a certain proportional speed (i.e. doesn't account for hills and stuff)

Uses of Open-Loop (feed forward) Control

Ballistic Movements: Pouncing, reflex reaching and withdrawal, etc.

Reactive Control

Based on tight (feedback) loops connecting a robots sensors with its effectors Purely reactive systems do not use any representations or looking ahead

Why do we use 6 legged robots?

Because at any given time, they will have 3 legs on the ground, and thus always a polygon of stability. With 4 legs, we could only ever move one leg at a time and still have the polygon

Why do we need path planning?

Because with just closed-loop control (jacobian-based), we get things like: Joint limits, obstacles, self-collision, singularities

Limitations of Simple Encoders

Cannot determine direction since the binary waveform is identical for both directions of rotation

Sensor Fusion

Combining different sensors to great effect

What is a Turing Universal Language

Contains: sequencing, conditional branching, iteration It can compute the entire class of computable functions

Arbitration

Deciding between two mutually exclusive, but still triggered, conditions for a reactive robot We need a hierarchy or fusion of rules or learning of rules at run time

How much each of the architectures look ahead

Deliberative: Only look ahead - plan then execute Reactive: No look ahead - only react Hybrid: Look ahead with the brain - react quickly with the wheels Behavior-Based: Look ahead only while acting

Closed-Loop (feedback) control

Getting a system to achieve and maintain a desired state by continuously feeding back the current state and comparing it to the desired state, then adjusting the current state to minimize the difference

Efficiencies of DC Motors

Good: 90% Cheap: 50%

State vs. Representation

In principle, any internal state is a form of representation - what matters to architecture is the form and function of the representation Conventionally, Representation refers to manipulable models of the external world

Where is path planning usually performed?

In the configuration (joint) space. One example of path planning is the Rapidly-Exploring Random Trees

Some reasons for perception

Interaction, Quality Control, Surveilance

Is an optosensor an emitter or detector?

It contains both actually.

Sonar is based on time-of-flight. Why use high frequency sound?

It's less noisy

Why is an optosensor usually in IR?

It's less noisy

A robot acts through

Its ACTUATORS (e.g. motors) which drive EFFECTORS (e.g. wheels)

Ladar vs.. Sonar

Lazers have a better angular resolution, higher speed means higher sampling rate, shorter wavelength means fewer specular reflections But, less coverage, they're expensive, and they're heavier

What is one alternative to path planning for humanoid robots?

Learning "motion primitives" from human examples. E.g. the tennis playing robot

Steps in Visual Processing

Low-Level Processing: Smoothing, edge detection High-Level Processing: Scene Reconstruction (what world produced this image), Object recognition

A free body in space has how many degrees of freedom

SIX 3 are translational (x,y,z) 3 are rotational (roll, pitch, yaw)

What can rotary shaft encoders measure?

Shaft velocity, shaft position, shaft rotation count

Advantages of DC Motor

Simple, Cheap, Various Sizes, Easy to Interface

Effector

Something that actually does something in the environment A claw, etc.

Sonar bounces off of stuff. Explain Specularity vs. Diffusion

Specularity: Angle of incidence = angle of reflection Diffusion: Energy is absorbed and re-emitted at a broad range of angles

Internal State

State of the robot Sensed using the robot's internal state. Can be Stored/Remembered Velocity, mood, etc.

External State

State of the world Sensed using the robot's sensors

Criteria for Selection of Architectures

Support for Parallelism (of processes/behaviors) Hardware Targetability (how well can it be mapped to robot sensors, how well can the computation be mapped onto real processing) Run-Time Flexibility: Is run-time adjustment and reconfiguration (learning) possible and facilitated? Modularity: How is encapsulation of control handled, etc. Niche Targetability: how well can the architecture allow the robot to deal with its environment Robustness: How well does the architecture perform if individual components fail Ease of use Performance

Guiding Principles of Subsumptive Architecture

Systems are built from the bottom up If layer 1 fails, layer 0 is unaffected, but layer 0 can be affected by layer 1 Lowest layers handle most basic tasks Each component provides and does not disrupt a tight coupling between sensing and action No internal models

Actuator

The actual mechanism that enables the effector the execute an action Eg. Electric motors, hydraulic/pneumatic cylinders, etc.

Sense-Plan-Act

The original approach to control - inherently sequential (horizontal) - produces deliberative architectures Subsumptive Architectures (vertical) are the alternative

Continuous vs. Discrete Stuff

The real world is continuous, while the digital world is discrete. Mapping a continuous function to a discrete one is called sampling Mapping a continuous variable to a discrete one is called quantization

Zero/Non-zero error

The simplest type of error: simply measures whether the current state = desired state or not

Are there legs in nature?

Yes- sperm. Sperm spin their tails with little wheels.