York Jong's List: 5 Star

• The processing power and memory  capacity necessary to match general intellectual performance  of the human brain are estimated. Based on extrapolation of  past trends and on examination of technologies under  development, it is predicted that the required hardware will  be available in cheap machines in the 2020s.
• Programs need memory as well as processing speed to do  their work. The ratio of memory to speed has remained  constant during computing history.
• Dividing memory  by speed defines a "time constant," roughly how  long it takes the computer to run once through its  memory. One megabyte per MIPS gives one second, a nice  human interval.
• Machines with less memory for their  speed, typically new models, seem fast, but unnecessarily  limited to small programs. Models with more memory for  their speed, often ones reaching the end of their run,  can handle larger programs, but unpleasantly slowly.
• Customers maintain the ratio by asking  "would the next dollar be better spent on more speed  or more memory?"
• The megabyte/MIPS ratio seems to hold  for nervous systems too! The contingency is the other way  around: computers are configured to interact at human  time scales, and robots interacting with humans seem also  to be best at that ratio.
• faster  machines, for instance audio and video processors and  controllers of high-performance aircraft, have many MIPS  for each megabyte. Very slow machines, for instance  time-lapse security cameras and automatic data libraries,  store many megabytes for each of their MIPS.
• Flying  insects seem to be a few times faster than humans, so may  have more MIPS than megabytes. As in animals, cells in  plants signal one other electrochemically and  enzymatically. Some plant cells seem specialized for  communication, though apparently not as extremely as  animal neurons. One day we may find that plants remember  much, but process it slowly
• With our conversions, a 100-MIPS robot, for instance  Navlab, has mental power similar to a 100,000-neuron  housefly. The following figure rates various entities.
  
• 
    MIPS and Megabytes.
• Universal computers  can imitate other entities at their location in the  diagram, but the more specialized entities cannot. A  100-million-MIPS computer may be programmed not only to  think like a human, but also to imitate other  similarly-sized computers. But humans cannot imitate  100-million-MIPS computers--our general-purpose  calculation ability is under a millionth of a MIPS.
• Computers doubled in capacity every two years after the war, a pace that became an industry given: companies that wished to grow sought to exceed it, companies that failed to keep up lost business. In the 1980s the doubling time contracted to 18 months, and computer performance in the late 1990s seems to be doubling every 12 months.
• 
   Faster than Exponential Growth in Computing Power. The number of MIPS in $1000 of computer from 1900 to the present. Steady improvements in mechanical and electromechanical calculators before World War II had increased the speed of calculation a thousandfold over manual methods from 1900 to 1940. The pace quickened with the appearance of electronic computers during the war, and 1940 to 1980 saw a millionfold increase. The pace has been even quicker since then, a pace which would make humanlike robots possible before the middle of the next century. The vertical scale is logarithmic, the major divisions represent thousandfold increases in computer performance. Exponential growth would show as a straight line, the upward curve indicates faster than exponential growth, or, equivalently, an accelerating rate of innovation. The reduced spread of the data in the 1990s is probably the result of intensified competition: underperforming machines are more rapidly squeezed out. The numerical data for this power curve are presented in the appendix.
• Chip progress not only continued, it sped up. Shorter-wavelength light was substituted, a more precise way of implanting impurities was devised, voltages were reduced, better insulators, shielding designs, more efficient transistor designs, better heat sinks, denser pin patterns and non-radioactive packaging materials were found. Where there is sufficient financial incentive, there is a way. In fact, solutions had been waiting in research labs for years, barely noticed by the engineers in the field, who were perfecting established processes, and worrying in print as those ran out of steam. As the need became acute, enormous resources were redirected to draft laboratory possibilities into production realities.
• The wave-like nature of matter at very small scales is a problem for conventional transistors, which depend on the smooth flow of masses of electrons. But, it is a property exploited by a radical new class of components known as single-electron transistors and quantum dots, which work by the interference of electron waves. These new devices work better as they grow smaller. At the scale of today's circuits, the interference patterns are so fine that it takes only a little heat energy to bump electrons from crest to crest, scrambling their operation. Thus, these circuits have been demonstrated mostly at a few degrees above absolute zero. But, as the devices are reduced, the interference patterns widen, and it takes ever larger energy to disrupt them. Scaled to about 0.01 micrometers, quantum interference switching works at room temperature.
• 
   The big freeze. From 1960 to 1990 the cost of computers used in AI research declined, as their numbers dilution absorbed computer-efficiency gains during the period, and the power available to individual AI programs remained almost unchanged at 1 MIPS, barely insect power. AI computer cost bottomed in 1990, and since then power has doubled yearly, to several hundred MIPS by 1998. The major visible exception is computer chess (shown by a progression of knights), whose prestige lured the resources of major computer companies and the talents of programmers and machine designers. Exceptions also exist in less public competitions, like petroleum exploration and intelligence gathering, whose high return on investment gave them regular access to the largest computers.
• 
   
   Agony to ecstasy. In forty years, computer chess progressed from the lowest depth to the highest peak of human chess performance. It took a handful of good ideas, culled by trial and error from a larger number of possibilities, an accumulation of previously evaluated game openings and endings, good adjustment of position scores, and especially a ten-million-fold increase in the number of alternative move sequences the machines can explore. Note that chess machines reached world champion performance as their (specialized) processing power reached about 1/30 human, by our brain to computer measure. Since it is plausible that Garry Kasparov (but hardly anyone else) can apply his brainpower to the problems of chess with an efficiency of 1/30, the result supports that retina-based extrapolation. In coming decades, as general-purpose computer power grows beyond Deep Blue's specialized strength, machines will begin to match humans in more common skills.

15 more annotations...

• This paper describes how the performance of AI machines
  tends to improve at the same pace that AI researchers get
  access to faster hardware. The processing power and memory
  capacity necessary to match general intellectual performance
  of the human brain are estimated. Based on extrapolation of
  past trends and on examination of technologies under
  development, it is predicted that the required hardware will
  be available in cheap machines in the 2020s. - York Jong on 2007-09-09

• 由于AVR的Flash存贮器可方便的使用ISP技术在线的多次擦写，因此建议尽量不使用（依赖）仿真器来开发和调试程序。
• 尽量使用高级语言编写系统程序。
• 使用Atmel公司提供AVR Studio软件模拟仿真环境，以及其他的软件模拟仿真环境（BASCOM-AVR）。
• 尽量合理和充分使用AVR片内的资源，如EEPROM、A/D、内部的RC振荡源。
• 尽量采用串口通信连接的外围器件，大容量的存储器、LCD控制器、打印机、不用8279（LED数码管+键盘）而使用7279等。除了必须外扩RAM（如语音和图象），一般不提倡使用并行扩展（573+译码电路），减小硬件和连线以及PCB板上错误的出现概率，同时也提高了系统的可靠性。并行扩展向串行扩展是发展趋势。现在有大量的新的外围器件采用高速的串行接口，如A/D、D/A、RTC、存储器等。
• 尽量使用以及在目标板上预留ISP程序下载接口，或使用IAP技术。
• ISP接口与I/O的兼容性比JETAG好。
• 注意和掌握AVR配置熔丝位的使用

• 提高硬件可靠性的考虑:

   

   

  
    .尽量采用片内晶体、采用低频率的系统时钟、振荡电路的输出小幅度。
   
    .选择合适的启动延时参数
   
    .使用BOD、片内的看门狗
   
    .合理休眠方式的使用
   
    .不用I/O口设定输出低电平
   
    .利用内部的EEProm和寄存器MCUCSR判断复位标志，进行不同的处理
• 嵌入式系统目前以大量的使用串行接口外围芯片和各种通信接口，如RS232、两线（I2C）、三线（SPI）、单总线、USB、CAN、TCP/IP等。开发人员和程序员应了解低层协议，熟悉硬件怎样和如何实现低层协议，如何定义可靠的上层应用协议，以及低层协议驱动同上层应用协议之间的接口设计（中间层软件的实现）等。

• 通信接口的编写应尽量

   

   

  
    .采用中断+缓冲区，
   
    .分层+结构化设计，
   
    .尽量不使用轮循方式(降低AVR的效率)。参见URAT（RS232）驱动+中间层软件示例。
• 尽量不使用传统的前后台（中断）系统设计模式，任务之间相互影响和干扰，无法定时操作。
• 采用TimeTip+状态机设计+CASE结构，实现多任务并行运行系统设计方法。或时间触发式的系统设计。
• 移植小型嵌入式操作系统

12 more annotations...

• Many Logo-based and other robotics products produced by LEGO   are distributed to schools in the USA by Pitsco.
• This free-range turtle does not require connection to a computer.   All the controls are on board.
• The Cricket is a tiny computer, suitable for all kinds of   robotics projects, that you can program using Logo.

1 more annotation...

• Logo has long been used to control mechanical
  turtles and other robotic devices. Here are some sources of
  equipment and related software. - York Jong on 2007-06-17

• • procedure definition with inputs and return values;   
  • global variables and local procedure inputs;   
  • control structures like if, repeat, and  loop;   
  • a 16-bit number system (addition, subtraction, multiplication,  division, remainder, comparison, bitwise operations, random  function);   
  • motor and sensor primitives;   
  • timing functions and tone-playing functions;   
  • data recording and playback primitives;  
  • communications primitives.
• 
• The text area in the lower left is the “Command Center.” Statements typed into this window are immediately transmitted to the Cricket and then executed.
• Motor commands are used by first selecting the motor (using a, b, or ab,) and then telling it what to do

• ab, onfor 20 

    

  There is also a free-running timer, which keeps track of elapsed time even when the Cricket is doing other things.

• global [cats dogs] 

    

  creates two globals, named cats and dogs. Additionally, two global-setting primitives, setcats and setdogs, are instantiated.
• There is a single global array for storing data which holds 2500 numbers.
• To transfer data to the PC, use the standard Cricket Logo software. In the “Cricket” menu, there is an option “Upload Data.” This will bring up a dialog box that allows you to retrieve the data from the Cricket and then save it into a text file.
• Please note that the memory for keeping track of recursive calls during execution is quite small, and limited to perhaps six calls deep. If the stack overflows, the Cricket halts and beeps five times.
• The Cricket supports tail recursion, converting the recursive call into a goto statement, and avoiding the stack depth problem just mentioned.
• Cricket can have one “background task.” This task repeatedly checks a condition. When this condition becomes true, the task interrupts foreground activity and processes its special action. When the background task's action finishes, execution picks up where it left off in the foreground activity.
• The action executes exactly once each time the condition goes from false to true

10 more annotations...

• The Handy Cricket is programmed in a language called “Cricket
  Logo,” which is a simplified version of the powerful yet
  easy-to-learn Logo language. - York Jong on 2007-06-16

• All Cricket devices have a built-in bidirectional infrared communications channel, which is used for Cricket-to-desktop communication (when downloading programs to a Cricket, or viewing sensor data) and Cricket-to-Cricket communication.
• Cricket Logo is based on an iterative, interactive model of project development. It includes a “command center” window; instructions typed into this window are instantaneously compiled, downloaded to a Cricket, and executed, giving the system the flavor of an interpreted software environment such as LISP, BASIC, or FORTH.
• The MetaCricket software system is based on a virtual machine, written in PIC assembly language and running on the Cricket, and a compiler for the virtual machine running on a desktop development computer
• It is straightforward to implement an interpreter-like interface, where user commands are transparently compiled, downloaded, and executed.
• The Cricket virtual machine is burned into the PIC microprocessor's internal ROM
• The user's code resides in a serial EEPROM
• Built-in infrared communications routines include a protocol for reading and writing to this external EEPROM, and for asking the virtual machine to begin execution of byte codes already loaded into the EEPROM.
• User-level primitive functions compile to one, two, or three bytes of object code for the Cricket virtual machine.
• The Cricket virtual machine has two process threads: a foreground process and a background daemon. In most Cricket programs, the foreground thread handles all the work, but for some tasks, the background daemon is valuable. For example, the background daemon can be used to instigate a periodic activity, or take action when some event occurs.
• There are hardware-specific primitives for interacting with on-board Cricket hardware. Motor commands set state (on or off), direction, and power levels for each of the two integrated motor drivers. Analog sensor primitives (sensora and sensorb) return a value (0 to 255) for each of the two voltage inputs. These inputs also may be interpreted as digital values using the switcha and switchb primitives. There is a pair of primitive functions for generating tones on the piezo beeper: beep and note, the latter taking pitch and duration arguments.
• there is a background millisecond timer that is updated every four milliseconds
• One foreground thread plus one background daemon   Daemon fires when provided Boolean expression makes false-to-true transition
• Users write programs for the Cricket in Cricket Logo, a dialect of Logo specialized for the Cricket virtual machine. Essentially, there is a one-to-one mapping between statements in Cricket Logo and primitive functions built into the virtual machine. This makes the implementation of the compiler far simpler than typical compilers.
• The compiler includes an interactive mode—a text window where user expressions are compiled, downloaded, and executed in one step when the user presses the return key. A portion of the Cricket's memory is set aside for these dynamic programs.
• we have found that a debugger is not necessary because of the interactive and incremental style of project development that occurs when using the Cricket.

•   The infrared protocol includes the following capabilities:  

    

  •   Check that a Cricket is present and ready for other commands.  
  •  Write a byte to the Cricket's EEPROM.  
  •  Read a byte from the Cricket's memory.  
  •  Begin program execution from a particular memory address.

14 more annotations...

• 
• Suppose you have a sensor that can see many different colors, such as a photoresistor. How would you use this sensor to detect red apples vs. green apples?
• If you shine a red light (such as from a red LED) on both apples, the red apple will reflect much more light than the green apple. As such, the apple that appears the brightest to your sensor will be the red apple.
• photoresistors have different levels of sensitivity to different colors. Typically photoresistors are optimal and peak at around 520nm. They are also sensitive to distance away from the target.
• 
• The ambient light (for example, sunlight coming in through your window) can change throughout the day. You will need to shield your sensors from this light.
• different color LED's shine at different wavelengths and different brightnesses,
• 
• In reality you must calibrate your sensors before they will work.

• equation: 
   

   abs(new_reading - calibrating_reading)/calibrated_reading * 100 = similarity  

   now using our numbers: 
   grey floor = (108 - 95)/95 * 100 = 13.7% different 
   white line = (108 - 112)/112 * 100= 3.6% different  

   compare: white line < grey floor 
   therefore the sensor sees a white line
• To detect two colors, of an infinite number of shades, you just need one LED and one sensor.
• 
• 

• if you wanted three different colors it would look something like this:  

   
• Programming your color sensor is very easy. You simply turn on a LED with a digital output pin, wait about 50ms for the photoresistor to change (photoresistors are much much slower than infrared sensors), take a reading with an analog pin, then turn off the LED (if you have more than one LED).
• 
• your color range sensor won't have a range of more than a few inches at most, depending on the brightness of your LED.
• So what if we had the LED shine for only 50ms? Well, that's 1 / .05 = 20, or 1/20th of a second. This means that if you shine your LED for only 50ms, then it can take about 10 to 20 times more current, and therefore shine about 10 to 20 times brighter.

16 more annotations...

• 
  Suppose you have a sensor that can see many different colors, such as a
  photoresistor. How would you use this sensor to detect
  red apples vs. green apples? - York Jong on 2007-06-14

• 
• The Infrared emitter detector circuit is very useful if you plan to make a line following robot, or a robot with basic object or obstacle detection.
• Place the LED and detector an 1/8 to a 1/4 inch apart maximum, basically parallel and almost contacting.
• Aluminum foil covered with electrical tape works as a great shielding material.
• Many emitters are strong enough that if you aim the robot at a white wall, and turn off all other lights in the room, you can see how the IR light is projected from your robot. Good debugging aid.
• Output your IR values on your computer screen real time to optimize positioning and Pot calibration (depends on if you want range detection or white line detection).
• dark colors absorb more light than clear, so the dark LED is the detector
• The general rule of thumb with sunlight shielding is if you cannot see any data reading difference inside or outside, your sheilding is effective enough to work. Film canisters or electrical tape works very well. A modulated signal (such as in remote controls) also reduces external interference.
• Depending on resistor values, your IR circuit can be tweaked to better detect color instead of distance

7 more annotations...

• 
  The Infrared emitter detector circuit is very useful if you plan to make a
  line
  following robot, or a robot with basic object or obstacle detection. - York Jong on 2007-06-14

• how to interpret sensor data into a mathematical form readable by computers

• There are only 3 steps you need to follow: 

   

  • Gather Sensor Data (data logging)
   
  • Graph Sensor Data
   
  • Generate Line Equation
• Some sensors (such as sonar and Sharp IR) do not work properly at very close range
• The way to get rid of noise is get a bunch of readings, then only keep the average. Make sure you test for noise in the actual environment your robot will be in

2 more annotations...

• 
  Most roboticists understand faily well how sensors work. They understand that most
  sensors give continuous readings over a particular range. Most usually understand
  the physics behind them as well, such as speed of sound for sonar or sun interference
  for IR. Yet most do not understand how to interpret sensor data into a mathematical
  form readable by computers. - York Jong on 2007-06-14

• 
  List of robot sensors, with detailed information and uses. - York Jong on 2007-06-14

• 镍镉电池所产生的电流大到足以熔化金属和引起火灾。其他电池只能产生弱电流
• 电流中所有化学和机械因素的净效应可表示为单一数学因数一等效内阻。降低内阻可得到较大的电流
• 加恒流(0.05C-1C)对NiCd电池充电。某些低成本充电器借助绝对温度终止充电。虽然简单、成本较低，但这种充电终止方法是不精确的。较好的方法是用检测电压跌落的方法终止充电。-△V 方法对于充电速率为0.5C或更高速率的NiCd电池是最有效的。-△V充电终止检测应该与电池温度测量相结合，因为变质电池和失配电池可降低△V
• 可以用检测温度增加速率(dT/dt)实现更精确的满充检测，这种充电检测方法比固定温度终止方法更好。基于dT/dt和-△V终止组合方法的充电终止方法具有较长的寿命周期，可避免过充电
• 因为NiCd电池的电荷接收度接近100%，所以在开始70%充电期间吸收几乎所有的能量，而电池保持微冷。超快速充电器利用该特点，在几分钟内把电池充电到70%电平，所加电流等于几倍的C率，而无热量产生。达到70%电平之后，电池以较低速率继续充电，直到电池充满电为止。最后，加0.02~0.1C涓流结束电池充电
• 尽管NiMH充电器与NiCd充电器类似，但是，NiMH充电器采用dT/dt方法，这是NiMH电池充电的最好方法。NiMH电池的充电结束电压下降比较小，而对小充电率(低于0.5C，这取决于温度)可以完全无电压下降
• 不能用NiCd充电器为NiMH电池充电，除非它是用dT/dt方法终止充电。因为NiMH电池不能吸收过充电，所以，涓流充电必须比NiCd小(0.05C左右)
• 慢充电NiMH电池比较困难，这是与0.1C-0.3C范围C率有关的电压和温度分布不能提供足够精确充满电状态的指示。因此，慢充电器必须依靠定时器来指示何时充电周期应该结束。所以，为了使NiMH电池充满电，应该施加接近1C(或根据电池制造商标定的C率)的快速充电，同时监控电压(△V=0)和温度(dT/dt)来确定何时充电应该结束
• 过充电(或过放电)锂离子电池可能会导致电池爆炸和人员伤害，所以，在使用这类电池时，安全是主要关心的问题。因此，商用锂离子电池组，包含象DS2720这样的保护电路(图7)，DS2720提供可充电Li+电池应用所需的所有电池保护功能：充电期间保护电池，防止超量电流流过的保护电路和限制电池耗尽，电平使电池寿命最长
• NiMH充电器可适应NiCd电池，但反之不行。专用于NiCd电池的充电器将会过充电NiMH电池。快速充电可增强镍基电池的寿命和性能，这是因为快速充电降低了内部结晶所引起的记忆效应。镍和锂基电池要求不同的充电算法。Li+电池需要保护电路来监控和保护过流，短路，过压和欠压以及过温

8 more annotations...

• Think hard, then act
• Don't think, (re)act
• Think and act independently, in parallel
• Think the way you act.
• • Thinking is slow. 
  • Reaction must be fast. 
  • Thinking allows looking ahead (planning) to avoid bad actions. 
  • Thinking too long can be dangerous (e.g., falling off a cliff,  being run over).  
  • To think, the robot needs (a lot of) accurate information. 
  • The world keeps changing as the robot is thinking, so the slower  it thinks, the more inaccurate its solutions.
• Reactive control is a technique for tightly coupling sensory inputs and effector outputs, to allow the robot to respond very quickly to changing and unstructured environments. Think of reactive control as as "stimulus-response".
• In deliberative control, the robot takes all of the available sensory information, and all of the internally stored knowledge it has, and it things ("reasons") about it in order to create a plan of action
• In hybrid control, the goal is to combine the best of both Reactive and Deliberative control. In it, one part of the robot's "brain" plans, while another deals with immediate reaction
• This requires a "third" part of the robot brain, and as a result these systems are often called "three-layer systems."
• Behavior-Based control is based on inspiration from biology, and tries to model how animal brains may deal with hard problems of both thinking and acting. Behavior-based systems, like hybrid systems, also have different "parts" or layers, but unlike hybrid systems, they are not as different from each other. All of them are encoded as behaviors, processes that take inputs and send outputs to each other, quite quickly. So if a robot needs to plan ahead, it does so in a network of behaviors which talk to each other and send information around, rather than a single planner, as with hybrid systems.

8 more annotations...

• attach your Pager Motors to your Popper using two Fuse Clips, two Small Paper Clips, and no solder
• 
• 

• Set one straightened paper clip aside, you will use it at the end. Bend the two tips of one of the two paper clips as shown.

   

  

• Put it in through the fuse clip like this, but make sure the notch in the fuse clip is facing out. (The clip has one edge bent inwards. This is the part that has to face outwards). Study this next picture closely.

   

  

• From the position above, bend the paper clip up and then around the lead of the fuse clip as in the next picture.

   

  

• Bend the paper clip under the fuse clip...

   

  

• then up and over the fuse clip:

   

  

• then around its other lead and you're almost done with the first fuse clip.

    

  

• First clip -- wire is on the RIGHT side of the fuse clip leads...

   

  

    

  Second Clip -- wire is on the LEFT side of the fuse clip leads... But remember to make sure the notch in the fuse clip is facing out. (The clip has one edge bent inwards. This is the part that has to face outwards). 

   

  

   

  Follow all the steps above with the second clip and you get this:

   

  

8 more annotations...

• The purpose of a solar engine is to act like a power  "savings account" -- a small trickle of incoming energy is  saved up until a useable amount is stored
• A solar-powered robot can be made to work, even in  relatively-low light levels
• Solar cell size is minimized
• by far the predominant
• theoretically the  most efficient

3 more annotations...

• At the heart of most solar-powered robots is a circuit called the solar engine (variously called Solar Engines, solarengines, SEs; a.k.a,relaxation oscillators). The purpose of a solar engine is to act like a power "savings account" -- a small trickle of incoming energy is saved up until a useable amount is stored. This stored energy is then released in a burst, in order to drive some useful (if only sporadic and incremental) work. - York Jong on 2007-05-28

•   >

     

•   >

     
• By just tying a neuron's  bias resistor to Vcc, rather than to ground, you can make a  "regular logic" (active high) Nv:
•    
• putting diodes and other resistors in parallel to give  different charge vs. discharge rates

•    

4 more annotations...

• One essential difference is that the Nv  responds immediately to an input, and sends the output for a  time duration -- the delay occurs AFTER the output is sent.  The Nu  responds to an input after a delay and sends the output  continuously -- the delay occurs BEFORE the output is  sent.
• "on" first, then a delay, then  "off"
• delay, then "on", stays "on"

1 more annotation...

• We use them as 3- or  5-volt triggers
• 1381s are CMOS  voltage-controlled triggers -- these "gate" a source until  the voltage is above some "trip" limit, at which point it is  allowed onto a third pin
• If you can't find 1381s locally, you might have better  luck finding its European cousin, the TC-54 -- for details  on it
• 
• any *cores built with a 74*04 will require additional logic  "downstream" to amplify the current  to levels sufficient to drive a moto
• 
• This chip is often considered the heart of Nv  net technology
• Schmitt  triggers can't easily be used in suspended  bicore implementations
• use its buffers as little current  amplifiers
• 
• The '240 is often called "the bicore  chip," because we can take advantage of the 240's inverters  to turn a single 74*240 into a bicore
• The '240 also has  tri-state outputs, so an enable line can be used to turn its  outputs on and off simply (good for adding reversing  capability to a 'bot).
• it is usable for either grounded  or suspended  bicore  designs (but better for suspended)
• 
• The '244 provides us with 8 (thus the "octal") buffers,  enableable in banks of 4. This is a very useful chip for  amplifying small currents
• it can drive up to 4 motors in 2 directions  each, or you can "buddy up" inputs and outputs to drive  fewer motors at higher current
• 
• The '245 is an octal buffer  chip, and so has 8 channels of buffering  power available for our misuse. This chip was designed for  data transmission uses, but we'll misuse it as a motor  driver chip
• it can drive up to 4 motors  in 2 directions each, or you can "buddy up" inputs and  outputs to drive fewer motors at higher current
• 
• Note that if you need more than  about 200 mA per motor, you'll need to use an H-bridge, or  some similar motor  driver
• 74HC/HCTxx non-buffers  (74HC14 or 74HC04) draw about half of the current  consumption, and have about half the drive current  compared to HC / HCT buffer  chips (74HC240 or 74HC245). Non-buffer  chips are thus better for oscillators, say Nv  and Nu  applications; they are not suited for use in driving  motors.
• The ideal BEAM  circuit would use a low (2V-3V) voltage core and sensors  combined with level shifting high (5-6V) volt motor  drivers to maximize efficiency.
• 74ACxxx used in typical BEAM applications uses 4x more  supply current  than does 74HC/HCTxxx.
• 74AC is best suited for motor driver  applications with all inputs driven rail to rail.

23 more annotations...

• The following material is intended to cover usage and part selection details of ICs you're most likely to see in BEAM robots.
  - York Jong on 2007-05-27