TOMI™ Technology in Small Automatons
Russell Fish III
Automatons are machines that perform specific tasks without human intervention. An automaton must possess the following attributes:
- An established measurable goal,
- A means of being aware of the physical world,
- Intelligence and the ability to make decisions based on its awareness,
- A means of interacting with the physical world based on the decisions,
- A means of learning from the results of its actions, comparing to the goal, and adjusting its behavior.
Small automatons will range from marble size to volleyball size. Cost will range from $10 to $200.
In addition to being smart, small automatons will have the ability to coordinate interaction with other automatons to achieve a goal.
An automaton can be thought of as a complex feedback system. A simple feedback system must have a goal to seek. An engine cruise control will seek the speed set by the motorist. A thermostat will seek the temperature set by the homeowner. A more complex system can seek multiple goals. Human behavior consists of seeking food, water, warmth, and a good job to pay for the sports car. In humans the drive toward these goals is called motivation.
A simple feedback system can be implemented by explicitly specifying the steps necessary to seek the goal. For the cruise control example:
1. Measure speed
2. If speed is too high, reduce fuel flow. Otherwise, increase fuel flow.
3. Rinse and repeat.
Complex systems require a different strategy, particularly when goals may conflict. For the human example:
1. Measure money in wallet.
2. Measure hunger. Measure thirst. Measure temperature discomfort. Measure car desire.
3. Good grief. What to do now?
Goal seeking behavior is a particularly efficient problem solving approach for multi-goal systems because rather than explicitly specifying the steps of a solution, the various goals can be quantified, weighed against each other, compared for temporal aspects (If I get the car today, will I starve tomorrow?), and take into account learned history.
Simpler goal sets can be satisfied with simpler automatons. One of the few commercially successful consumer automatons is the Roomba vacuum cleaner by iRobot. At its simplest, Roomba traverses a flat surface and changes directions when it encounters an obstacle while sweeping up dust. Roomba barely qualifies as an automaton, but definitely qualifies as a commercial success.
Future Small Automaton Goals
The following are some examples of goals that can be physically achieved if sensors, learning software, and cost limitations can be removed:
1. Build a house.
- Lay a foundation
- Build a frame
- Route utilities
- Apply flooring and wallboard
- Apply roofing
- Avoid running into people, pets, and the street.
2. Clean a sidewalk
- Identify gum and remove
- Identify light weight trash. Retrieve and store.
- Identify metal trash. Retrieve and store.
- Avoid running into people, pets, and the street.
3. Clear a minefield
- While maintaining a safe distance from other automatons, locate next mine.
- Report mine location to other automatons.
- Destroy self and mine.
- Avoid destroying people.
4. Harvest a crop
- Identify fruit and classify as ready for harvest.
- Remove fruit from tree or vine.
- Return harvest to central location when fully loaded.
- Avoid running into people, pets, and the street.
5. Exterminate insects
- Identify target insect.
- Determine whether alive or dead.
- If alive, terminate with prejudice (laser or physical disruption).
- Retrieve and store corpse.
- Repeat until power depleted or human commands to stop.
- Avoid terminating people.
Awareness for simple systems often consists of sensors; temperature, speed, pressure, etc. For humans, awareness consists of the 5 senses.
Automatons will have increasing sensors and senses with increasingly sophisticated goals. Existing automatons have the ability to sense; light, images, sound, temperature, acceleration, global position, and power status.
To achieve the goals above will probably require some of the following sensors:
1. Visual (or other electromagnetic) - Feature recognition is a key requirement for sophisticated automatons. The original TOMI technical requirement was to interpret images from the 5 Mars Lander cameras. The goals included avoiding obstacles and dangerous drop offs as well as identifying interesting rocks to probe.
Cameras and other optical sensors are a small component of visual awareness. Interpretation of the raw data into actionable intelligence is required. For instance, visual feature identification is useful. Raw photos are great for the grandkids but useless for automatons . Stereo vision offers a much broader insight into features but also more challenging computational burdens.
2. Acoustic - This is a less important sense than sight, but it can be useful for some goals. Flying insects can be efficiently located acoustically.
As with visual sensors, the raw acoustic data is of little use. Interpretation is required.
3. Tactile - A goal that requires physically manipulating objects can be made more effective with sensors that detect contact and pressure. As a simple example, the Roomba has bumper switches to detect obstacles. An automaton that is performing construction or picking fruit will be very difficult to implement without tactile feedback.
4. Position - Some people think automaton position is limited to establishing GPS coordinates. GPS provides an accurate absolute position.
However relative position is far more important for many operations. Consider a human driving a nail. The absolute GPS coordinates are highly unlikely to be accurate enough for successful driving. However the direct visual observation of the relative position of nail, hammer, and lumber provides the necessary information to successfully drive the nail through the lumber. The same is true for many automaton positioning operations.
For example, the automaton painting a house is better served by visual detection of the edge of the last paint application that the absolutely coordinates of the house, the wall, and the last known location of the paint head.
An automaton clearing a mine field need only know the relative location of itself to its peers and the nearest mine.
The human model of establishing position by visually identifying objects and landmarks is probably an efficient one for many automaton applications.
Once the sensors have captured the awareness information and interpreted it, the automaton must devise a course of action to achieve the goals. Sometimes intelligence will include one or more strategies. A strategy may be thought of as a rule of thumb to apply to a problem rather than an explicit step-by-step solution. A strategy is flexible because unlike a step-by-step solution it may be applied to never before encountered situations.
Watch a bird attempt to open a plastic wrapped item. The bird is likely to have never encountered such a challenge in the wild, yet he will use the same strategies that open nuts or captures beetles. Often the strategy will work even though it is applied in a totally new context.
Automatons will implement similar strategies.
Actuators are the best developed element of automatons since actuators with similar specifications have been directly controlled by humans for decades. Physical interaction can consist of moving the automaton, pushing objects, or grasping items. Special purpose physical interaction would include tools to destroy insects or devices to destroy the automaton along with a physically proximate mine.
Learning will be the key capability that leverages automatons into extensive use. Through learning an automaton increases its effectiveness over time as opposed to repeating the same mistakes. Learning allows the automatic creation of new strategies when faced with new and different circumstances.
The Effect of TOMI Technology on Automatons
TOMI Technology can enable automaton development in the following areas:
1. Sensor interpretation.
4. Power consumption
Small automaton sensors will likely be visual and tactile. They will also be acoustic, but not for distance determination. Acoustic sensors will be used for speech recognition to interface with people.
Visual and acoustic interpretation algorithms are similar. Mostly they are various approaches to pattern identification. The Hidden Markov Model (hmm) is one popular algorithm. Hmm maps well to the TOMI caches since its inner loops are spatially local for both instructions and data. Integer FFT also runs efficiently on TOMI.
Automaton intelligence will be less step-by-step than today's Roomba and more goal seeking and strategy exercising. The software is likely to resemble adaptive neural fuzzy logic. The knowledge base of such a system can often be factored across many nearly independent CPUs. Due to the independence, acceleration is nearly linear with additional CPUs.
The TOMI architecture is well suited for such a system because of the tightly bound relationship between CPU and memory and its efficient manner of implementing many CPUs on the same die enables a massively parallel system with reasonable cost and power consumption.
Automaton learned behavior is likely to be stored in flash RAM. Initial automatons will likely be implemented using discrete flash and TOMI chips. The learned behavior will be read into TOMI memory and operated on there. As automatons mature, it is likely that TOMI CPUs will be integrated into flash RAM in a manner similar to the integration into DRAM. The advantages of such a configuration will be a several order of magnitude increase in pattern matching speed.
All of the above capabilities can be accomplished by existing computing technology. However small automatons cannot be commercially successfully with the current state of the art microprocessors. They are too power hungry and too expensive.
JPL representatives originally approached us about TOMI due to their severe power budget. Small automatons must expend almost all their power on actuators. Not only is a 500Mhz TOMI only 23mw when running at full speed, but its Hibernate mode of 12Mhz is only 300uw. Automatons will require intense computation when sensors detect changes, but most of the time they will be loping along waiting for something to happen. This is an ideal operating environment for TOMI.
Finally, eventually everything comes down to cost. Small automatons will be expendable laborers, much like graduate students. The computing requirements are pretty rich with speech recognition, image identification, and neural logic. The computing budget, not so rich.
DRAM and flash devices are the cheapest transistors on earth with as much as a 100/1 advantage over logic devices. TOMI is built using DRAM or flash transistors. You cannot build a less expensive microprocessor, particularly one that offers the low power consumption and high computing power.