Boston Dynamics Robotics - Updates


Recommended Posts

On 2/2/2017 at 4:04 PM, SALSN said:

Very interesting, but can we please remember that these robots have next to nothing to do with genera purposel AI at the moment.
The worst thing a balancing algorithm in revolt will do is tip over the robot.

But I definitely agree that if these systems were ever married to any kind of general purpose AI we got to be very careful indeed!

 

Thank you for your input. What you have stated is true, for general purpose AI, but not for primitive autonomous decision making skills, which are used in many control system venues. A few of these robotic platforms are a lot more than just a few balancing algorithms, which were in use decades ago. The problem here is the context of the conversation a few of us were having and how you interpreted it. I, for one, am guilty, particularly on Neowin, of not going into too much detail as I leave that for other venues and keep things in "fun mode" here...keeps my sanity. This invariably leaves out a lot of key factors that I have not expressed properly. I have spent a very long time in engineering with control systems being of the utmost importance in this endeavour. Below are two "short novels", one non technical and one partially technical. I have placed it in this fashion to help the curious if they wander across this mess.
//
While back in the hospital on Thursday, I had a bit of time on my hands, read this thread and thought I would jot down some drivel. The following is my opinion only, form your own opinions and don't mind the spelling mistakes or coherency...I'm a bit pre-occupied with health issues at the moment.
//
First we must get rid of the Hollywood interpretation of artificial intelligence. Will we ever be able to mimic a human brain? IMHO, No, the complexities, of which we still don't understand, are overwhelming and why would we want to. The ability to compartmentalize for maximum efficiency in a designated field will be the trick. Artificial intelligence is being studied at numerous facilities and progress is being made. I will leave out the major advances and do an academic comparison. We use very primitive decision making abilities in a lot of venues today. I would treat this as choosing from a complex multivariable multiple choice exam. The order of difficulty, in the future, increases to answering an exam with a detailed written professional narrative. Next level is writing the post secondary exam you are taking. Next step is writing the book that you study from then finally winning a Nobel Prize in your field. Every step is reason for concern.  
//
//
I will attempt to put this into an engineering context. All one needs (to be real dangerous) is a very primitive form of decision making capability, which we have had for some time. Abilities increase when the number of decision blocks increase as well as chosen hiearchy. The problem is keeping multiple, high quality environmental inputs (real time and historic) to an acceptable size which will allow processing power to perform close to "real time" control abilities. This, is no easy feat but hard work by many intelligent individuals have made great strides in this endeavor.
//
The latest video, with "Handle", is basically click bait for investors, which is why Steve Jurvetson and Peter Diamandis were there, of note since they are savy investors with a very keen eye for future adaptions. In fact, I saw the first release, which was removed as Boston Dynamics requested a portion to be "blurred" or removed. The real star was the new and improved "Spot Mini" with it's new "primitive" mobility regime.
//
Context time for a thought experiment. Let's start with a box of parts...
//
1) Vision system, visible and infrared...done deal as I have worked on very capable older systems and todays level of technology is stunning in comparison.

 

2) Proximity detection....very advanced, active and passive, throw in gps integration. 

 

3) Motion detection...today's capabilities are very adavance.

 

4) Processing cababilities for non radiation hardened projects....We all know what is available to industry, suffice to say, more than adequate.

 

5) Communications abilities....very adavanced, multiple venues including various levels of jamming mitigation.

 

6) High speed actuators....readily available.

 

7) Transducers...displacement, velocity, acceleration, temperature, acoustic, atmospheric, gaseous, magnetic, all available with speed/precision. 

 

8) Small to medium size robotics platforms with reasonable mobility (doesn't need to be great)... Do I really need to state availability, widespread, but lets use the examples of "Spot Mini' and "Cheetah", which holds the speed record at 29 mph (try to outrun this one).

 

9) Available attachments that can do harm....Pick your own,sky's the limit but here are a few..tear gas dispenser, sharp blade, piercing probe, tazer, hardened projectile, rotary cutting device, how about razor sharp claws,... for the military, explosives, automatic weapon, grenades, target tagging for auxillary assault...etc.

 

10) Terrain attachments....pick a terrain and apply the type of robotic appendage ends to enable sure footed movement with some emphasis on agility.

 

11) Swarm intelligence....todays levels are primitive but display effective awareness for containment of purpose.

 

12) Primitive targeting system...No problem with a few tools above.

 

13) Extremely primitive artificial intelligence...already available, this can be described as a system which can monitor real time environmental sensor data live as well as from multiple updated sub sets and it will choose the most efficient method, to move it's actuators/apendages, in near real time, to guide the platform to an end task. If one needs convincing of this, one example is Lars Blackmores landing control systems, particularly for Red Dragon.
//
There are many more "tools" out there for the tool box, but lets clip it here for now.
//
 Location...pick a somewhat easy locale such as small warehouse room, gated off area or light scrub and grassy area with moderate hilly terrain.
//
Knowing what human nature is capable of, in particular, the military role for toys, let us choose 2 scenarios.
//
One scenario will be to contain, the other will be to eliminate the threat.
//
Now, one can throw a bunch of these "tools" on a robotic platform, in any level of complexity you choose.
//

For situational awareness, imagine one enters an area and a 60 lb dog, happy to see you, jumps up on you. Depending on your size, it could knock you over.

 

How about a 60 lb metal robotic dog doing the same...could hurt. How about a 200 lb. unit.

 

Now go into an enclosure of your choice, and have 20 of these units, loaded up with some "tools", in swarm behavior, with the sole purpose of your containment, pepper spray optional. Not fun.

 

Now go into an enclosure of your choice, and have 20 of these units, loaded up with all the "tools", in swarm behavior, with the sole purpose of your elimination. This is the problem. Even worse when we have half a dozen "function specific" robot sizes, including drones,  mixed in to spice it up.

 

Lets take it one more step. Let the military, with a nice budget, "improve efficiency" by using the latest available materials and the best "tools" to eliminate you. With todays available technology, an extremely primitive form of "ai" will be more than sufficient to do the job...which is why it is a nightmare.

 

I have been harping on this for quite some time, skip the scifi, it is inevitable for technology to progress and this requires early intervention for some form of "robotic laws". The problem is not just in the future, we have the problem now with very primitive "ai" because our sub systems have radically improved and can effectively change emphasis on requirements.

 

The real kicker. We are lucky that most terrorists are not playing with a full deck. If an intelligent bunch, with no morals, had the ability to unleash a portion of this destructive power .....

 

On a brighter note, it would be funny seeing Spot Mini, in a Tesla drivers seat, self driving to a computer shop, where Spot wanders in and brings out a laptop, purchased via paypal on the cellular system, on the way there.

 

One can skip this next (boring long winded) technical blurb and carry on to the end of this comment.

//

Here is a tech related example to describe a few parts of a very large and complex project, one employing primitive decision intelligence in near real time. We will start with a simple driver design, an amplifier subassembly which will drive a high speed actuator. We will use modern electrical engineering CAD for the final production design and analysis.This entails model and precision selection, topography style, placement, and production analysis.
//
We will then perform many prior modifications to the actual model due to the major contraint we have...time. Every clock cycle counts when compounded and takes us away from "near real time" operation.
//
An example being a single stage amplifier engineering model, descriptive to get the point portrayed. At it's most basic, the stage is a mathematical description, of the parameters involved in that stages operation. This is taken to the sub atomic level, where each one is described by a mathematical statement. A few examples are a leg current, ambient temperature, reverse currents, static charges, input noise, to name a few. Accuracy of the model depends on the level of detail given to each parameter and the number of parameters used, which can be a very large number. What am I getting at? Pick a FET or BJT, find the data sheet for it and look at the parameter listing given. One will generally see several pages along with characteristic curves nicely graphed for you.This is good enough for general use, but we have other plans. There are a few mathematical program constructs to develope stages, I personally use a variant of SPICE. We describe all parameters, many at the sub atomic level, by a mathematical statement (complex equation string). By increasing the "order" of the mathematical statement, we increase the processing time required to run the analysis's that we have chosen to run. The analysis describes the models overall activities, via the relations and dependancies of all parameters to each other. 
//
We are aware of the paradox used in physics, electron velocity or position, not both. Same applies here, accuracy or processing speed for your model.
//
Example...I can design a stage with a lot of low level parameters removed. The accuracy took a hit but I have graphed the major outputs of concern and know that I can compensate, off stage, by discreet autonomous loops, to handle what I gutted. What I want is the characteristics curves of importance. This analysis will run fast but lacks accuracy. Next, I crank up the parameters which results in an extremely accurate model which reflects a production stage, but it can take a long time. That is okay because I want to see the same characteristic output curves. I compare the two stage analysis methods. I look at what I deem to be the most important features, were they close enough where I can design effective off stage stabilities to keep the unit in a confined area of operation. Is the stage model I have chosen able to do what I want, or does a model change need to be considered. The end result is that I have mathematically modeled a very accurate representation of the physical model and know what I need to design, off stage, to keep the design focused purely on a few tasks.
//
The method above, or similar strategies, will be used to design individual discreet stages, programmable arrays and complex stage sub assemblies and the final completed project assemblies as an operational unit. It's a lot of work. Several software packages will be used to test electronics assemblies, mechanical assemblies and environmental physics.
//
The next project mini level will be how inputs to the designed stage are handled. I know how fast the basic stage is, but I want accuracy and speed for the overall system. The various stages have defined speed of operation and the actuator used will have known speed. The problem is the input. One direct input is fast to handle but contains little intelligence. We need transducer accuracy, fast sampling and a "history" data base as well as the most important...rates and trends..how is this data moving at this time, what associated variables from another subsystem can help us get a "heads up".This is a lot of data. Do we need all this data, or can we only focus on an important subset, depending on where we are on our operational curve. Can we even see the trend quickly. Now is where mathematical transforms come into play. A difficult equation to solve, with a lot of data, is time consuming and also complicates the physical design required to implement. But, if we use a transform, we may be able to spot a trend faster, design a stage with less parts and speed up a decision process. Switching between several mathematical domains allows us to not only reduce complexity, it allows us to see an association that we were not aware of which had a larger impact than we had assummed.
//
We now jump to the overall "corridor" of operation for the "machine" we are designing. This corridor sets up the area where we would like the machine to operate in. The corridor will be wide enough to allow "time for corrections" which enables us to remain in it, without getting too carried away with excessive control system manipulations. Preference will be for an autonomous mode of operation. The various areas of the corridor will have distinct parameters which are weighted. Outside the zone will require immediate action to bring the machine back into the operational area, or worse case, insert into another operational area due to unforseen environmental conditions. When in our "zone", just keep it there with minimal effort. As we wander from our zone, the response will have increased "attack" vectors to get us back into our comfy zone.
//
The decision process for the project's operation is critical. The sooner a decision is made, the sooner the corrective action is taken.The most important element is our real world data. The level and accuracy of the data that we obtain allows us to have a better understanding of our environment. With this understanding, we are able to refine what variables have importance, how to weigh this importance in a particular control scheme and decide on the scope of autonomous decision capabilities.
//
Finally, we have accurate real world environmental data. We have transducers that have the accuracy required for our operational needs. The speed is known for control surfaces/appendages as well as for the electronic subsystems. We decide on the corridor of operations and select the control schemes to be used for the various areas. We select the demarcation points where fixed reponses occur and where autonomous decisions can be made on it's own. We develope the methodology for autonomous thinking. This is determined by "where" in the corridor that the machine is in. What level of data will be presented and the associated weights. A profile will be chosen to follow, based on the data presented and the corrective abilities available.
//
In the end, we have a processing ability that acts like a fixed logic output or a primitive intelligence for a confined set of circumstances. The ability is achieved by use of multiple processing areas, using high speed detection algorithms to present past, present and projected data,for many parameters, in multiple fault tolerant systems. This data is presented to the "decision core". The area of corridor operations will determine parameter weights and the most efficient profile to follow. The decision is made, and the control systems carry out the profile requirements with continual moderation via the "intelligence" section.
//
In reality, we start from a concept. The concept may be centered around a very important part and the rest designed downward, to achieve our goals. We have the ability to design a very primitive intelligence module, one that can act in "near real time" by the use of distributed processing and creative mathematical analysis and modeling.
//
Hope that helped someone.

 

Closing comment, if you skipped the tech part
//
//
//
For near real time operations, processor speeds, by todays standards are fine. Timely data management is the issue.The problem is efficient extraction of multiple  relevant sets of data, from  large precision pools, which may have to be modified for very quick use at a specific moment in time. This is achieved by use of complex mathematical modeling. The model determines the design of the circuitry which can increase speed of operations and reduce circuit complexity to a point where programmable arrays can become major replacements for otherwise large circuit requirements.  
//
An extremely primitive "ai", which we have had for some time, can be extremely dangerous when coupled with todays state of the art "toys". It's the toys that cause the harm, the platform just uses the toy's to be somewhat aware of it's surrounding, to spot an "intrusion", then use it's mobility to get to the intrusion point and let other "toys" do their function. When done, the unit can go back to a charge port, top up, then leave and go on duty as fixed monitoring sentry or as a wandering one.
//
Again, picture a 1000 meter by 1000 meter gated area, at night with no lighting. Inside the compund are 20 to 30 units of various sizes and abilities such as high speed running, state of the art multisensor array for spotting intrusions, large heavier units carrying nasty toys, smaller agile units with a single weapon, some at rest but monitoring everything, others on mobile duty and all with the ability to be controlled from a central command facility or even from a "senior" robotic package that is one of the units in the compound, or in the air. Then include primitive swarm abilities on call.
//
Ask yourself an honest question. "Spot has control over his joystick" meaning all the robotic platforms in this scenario have very primitive "ai" abilities. The robotic platforms have been instructed to eliminated all threats. If one were to wander into this area, how long until robotic task completion. One would be foolish to underestimate the capabilities of these various units.
//
What I have talked about above would be equivalent to three words in a dictionary when it comes to control engineering. Do not underestimate technology, it has a habit of coming back to bite you. A case of worrying about a level of advancement, when the primitive version is just fine because the peripherals are the deciding edge.

 

If anyone read all this...I'll be amazed because I fell into naps typing it.  :D

  • Like 1
Link to comment
Share on other sites

Thanks for the extremely long comment I must admit, I only skim read it, but I think we are talking a bit past each other.
I don't dispute that a robot like this could harm someone, just that its AI is not advanced enough to hold any kind of ill intend, so any injuries inflicted by a robot like this would be purely by accident, or set in motion by an operator, never some kind of robot uprising.

So I'm afraid of them only in the way that I would be afraid of getting run over by a car.

 

Of course things could change, but why would anyone want a humanoid robot with advanced general AI? A platform like spot or petman would very likely be used for manual labour rather than any kind of advanced thinking, that could take place in a server with no arms or legs.

 

Anyway I'm super excited to get a personal robot butler one of these days :-P

Link to comment
Share on other sites

The Japanese are developing humanoid general AI to use them as in home caregivers for their rapidly growing senior population and handicappers, a bit of an improvement on capuchin monkeys, and as construction & factory assistants.  

 

Other aging nations would likely pile on, but many of these capabilities parallel infantry uses. Like a pilot having a UCAS fighter wingman, a few tactical humanoid AI's in a unit could do dangerous duties while using human equipment.

Link to comment
Share on other sites

Just now, DocM said:

The Japanese are developing humanoid general AI to use them as in home caregivers for their rapidly growing senior population and handicappers, a bit of an improvement on capuchin monkeys, and as construction & factory assistants.  

Reminds me of the movie; Bicentennial Man.

  • Like 1
Link to comment
Share on other sites

19 hours ago, SALSN said:

Thanks for the extremely long comment I must admit, I only skim read it, but I think we are talking a bit past each other.
I don't dispute that a robot like this could harm someone, just that its AI is not advanced enough to hold any kind of ill intend, so any injuries inflicted by a robot like this would be purely by accident, or set in motion by an operator, never some kind of robot uprising.

So I'm afraid of them only in the way that I would be afraid of getting run over by a car.

 

Of course things could change, but why would anyone want a humanoid robot with advanced general AI? A platform like spot or petman would very likely be used for manual labour rather than any kind of advanced thinking, that could take place in a server with no arms or legs.

 

Anyway I'm super excited to get a personal robot butler one of these days :-P

I had to laugh because you are quite right, our train of thoughts crossed. For that I owe you an apology. The other lesson of note, for me,  is not to jot notes when in the emergency room......

 

I would like to think that we are in agreement that "ai", as it stands today, is very primitive and we are a very long way from "Hollywood" depictions.

 

The above narrative, which I would have preferred to reword a bit, was intended to display the complexity and issues of control engineering, by giving a few examples , and show that we have already come a long way and do have the ability for primitive "ai", as it is in use today in various control systems.

 

The big issue to me is human nature. Control systems and advanced peripherals are at the level that a "morally deficient" individual/group will be able to cause the same damage, today, as one would with a slightly smarter "ai" in the future. The difference today, multiple advanced peripherals acting as inputs to a modern tracking system, which in turn act as a direction input for a mobile platform, which is more than able to get to the target, to carry out an instruction set. The instruction set will be a use of a weapon. The peripherals are the issue.

 

As a side note, I still want a "Spot Mini" with a T Rex head, for a pet.

 

:)

Link to comment
Share on other sites

On 5/2/2017 at 10:38 PM, DocM said:

The Japanese are developing humanoid general AI to use them as in home caregivers for their rapidly growing senior population and handicappers, a bit of an improvement on capuchin monkeys, and as construction & factory assistants.  

 

Other aging nations would likely pile on, but many of these capabilities parallel infantry uses. Like a pilot having a UCAS fighter wingman, a few tactical humanoid AI's in a unit could do dangerous duties while using human equipment.

There are multiple facets to what we understand by advanced AI, to have a robot fill a combat role in place of a person, or take care of the elderly, they obviously have to have quite advanced understanding of their surroundings and how to interact with them. But in terms of thinking for themselves and being smart like a scientist or a general commanding a war, they really don't need any of that. So In that manner they will still be way behind any kind of human intelligence and will never need this kind of intelligence to fill the previously mentioned roles. That is why I don't think these kind of AI will ever be anything to worry about.
 

 

On 6/2/2017 at 8:30 AM, Draggendrop said:

I had to laugh because you are quite right, our train of thoughts crossed. For that I owe you an apology. The other lesson of note, for me,  is not to jot notes when in the emergency room......

 

I would like to think that we are in agreement that "ai", as it stands today, is very primitive and we are a very long way from "Hollywood" depictions.

 

The above narrative, which I would have preferred to reword a bit, was intended to display the complexity and issues of control engineering, by giving a few examples , and show that we have already come a long way and do have the ability for primitive "ai", as it is in use today in various control systems.

 

The big issue to me is human nature. Control systems and advanced peripherals are at the level that a "morally deficient" individual/group will be able to cause the same damage, today, as one would with a slightly smarter "ai" in the future. The difference today, multiple advanced peripherals acting as inputs to a modern tracking system, which in turn act as a direction input for a mobile platform, which is more than able to get to the target, to carry out an instruction set. The instruction set will be a use of a weapon. The peripherals are the issue.

 

As a side note, I still want a "Spot Mini" with a T Rex head, for a pet.

 

:)

I think we largely agree as well, but I also think we are using some terms with different definitions. As a student of control engineering, I associate that term with PID controllers and state space models, controlling primarily psychical systems from motors and other actuators, to production lines, cooling systems and a robot's physical motion.
But you are using it to describe decision making on a more general level that could include moral questions as well, is that correct?

 

Link to comment
Share on other sites

On 2/6/2017 at 1:08 AM, DocM said:

Or I, Robot

The A.I. (2001) was a good one too.

 

The original intention for developing the AI/robots in films always appears to be a positive one, but then the story takes a u-turn.

 

In addition to the classics of the genre like the Blade Runner, the Terminator, Robocop, the Matrix, there are two relatively recent films that have really impressed me and I think these two deserve much more attention than they've received. Ironically, the titles are somewhat similar:

 

Ex Machina (2015)

The Machine (2013)

 

I also recommend the original Swedish TV-series called Äkta människor (Real Humans).

Link to comment
Share on other sites

On 2/7/2017 at 6:00 AM, SALSN said:

There are multiple facets to what we understand by advanced AI, to have a robot fill a combat role in place of a person, or take care of the elderly, they obviously have to have quite advanced understanding of their surroundings and how to interact with them. But in terms of thinking for themselves and being smart like a scientist or a general commanding a war, they really don't need any of that. So In that manner they will still be way behind any kind of human intelligence and will never need this kind of intelligence to fill the previously mentioned roles. That is why I don't think these kind of AI will ever be anything to worry about.
 

 

I think we largely agree as well, but I also think we are using some terms with different definitions. As a student of control engineering, I associate that term with PID controllers and state space models, controlling primarily psychical systems from motors and other actuators, to production lines, cooling systems and a robot's physical motion.
But you are using it to describe decision making on a more general level that could include moral questions as well, is that correct?

 

Excellent, you are immersed in control engineering and I did a long blurb for nothing...but then again it will hopefully assist those with an interest.

 

I agree with pretty much everything in the above comment. Think of "state space" being pushed to a whole new level where we also factor in baseline data as well as accumulated near present data to reference with real time transducer inputs, on the fly, to operate in near real time. This is really pushing state space and why transforms have to be modified for a particular use, and used where they would otherwise not be, ...because we need to build the system from the complexities of these mathematical models to operate in a corridor.

 

The biggest hurdle to all of this...proper selection of archived data, the format of and the presentation for a decision corridor, in real time. 

 

We have that ability now...and this platform will be able to harm a human. The machine has no morals, it is just following a complex decision process and executing the requirements of the programmer/operator, who is in charge of morals. The issue has always been the real time processing of environmental inputs. An example is ...do I reach around a bunch of objects to get a package or take the package by reaching over other obstacles. Do I walk around several objects or walk over top of them. If I run this fast and have to stop over there, when do I slow down. 

 

These are simple human thinking processes that we allow the machine to do. No "ai" soldier, it's a weapons package that thinks, primitive and acts with respect to input goals. One package can carry all peripherals or only a few and relay data to other units with different roles.

 

This is equivalent to the compound example. With the program requirements loaded, grab a chair and as an intruder breaches the fence line, the platform which can be on patrol, spots, targets and proceeds to the intruder to unleash something...probably not good for the intruder.

 

I really need to stop rambling....:D

Link to comment
Share on other sites

  • 3 weeks later...

They are a great collection...but, Spot Mini's future capabilities is the short range concern for me...:s

Link to comment
Share on other sites

  • 11 months later...
6 minutes ago, DocM said:

Should have consolidated their new videos long ago...

...and posted in the original thread, maybe. It's totally up to you thou.

Link to comment
Share on other sites

  • 2 weeks later...

It's amazing how far they've come from the first one. 

 

Also, every time I see them abuse one like that, I can't help thinking "I will remember this, Human."

Link to comment
Share on other sites

  • 2 months later...
  • 4 months later...

They have done well.

 

They don't need "ai" to be scary...just fast real time processing for predefined and "gray" GNC.

 

These should not be for sale ...too easy to be nefarious with these packages.

Link to comment
Share on other sites

This topic is now closed to further replies.
  • Recently Browsing   0 members

    • No registered users viewing this page.