Search This Blog

Friday, October 19, 2018

Mountain Driving the Tesla Model 3 - Ute Pass

Driving the Tesla in Ute Pass, from 8,500ft to about 6,000ft - and back using Autopilot and Cruise control features.  A snow-capped Pike's Peak at over 14,000 ft can be viewed among the clouds in the distance.  You can visit a variety of different climates in a single day in the mountains.
A follow-on from my prior blog post: At this point it's probably a good idea to review the various levels of autonomy for self driving vehicles as outlined by the Society of Automotive Engineers (SAE) - 6 levels from 0 to 5:
The Tesla adaptive cruise control is cited as being a level one feature, with the current version of Autopilot being a level two feature - though with future hardware and software upgrades, it appears they hope to achieve at least level 4 self-driving automation.  At the moment, it looks like their new AI chip will become available in early 2019.

At the time of  my initial testing, the current version of software for the vehicle was 8.1.  Of course as soon as I finished the initial testing, version 9 was announced.  As I write this version 9 is not only out, but has also been uploaded to my vehicle.  However, as I understand it, version 9 does not have any additional Autopilot features enabled for the driver - though some of the advanced features (Navigate on Autopilot) run in 'shadow mode' to help verify that the automation software is performing up to standards, while racking up miles on all of the vehicles that have the software installed.  In any case, I'll rerun some of the tests again on version 9 in the future to see if I see any major differences.

For the most part, the adaptive cruise control mode under optimum driving conditions in Ute pass appears to be relatively safe.  There are some exceptions though to keep in mind.

Using the cruise control to manage the vehicle's speed and distance from the car in front of in this last few week's testing has so far worked as expected - even bringing the vehicle to a complete stop if needed.  Heading down the pass seems to be somewhat better than driving up the pass.  As long as the driver is managing the driving and lane changes, it moved down the pass without a hitch.  However there was some concern regarding vehicles parked road-side that may need some additional attention.  The following three images show a vehicle parked on the shoulder, a truck in the same relative position (though moving, not parked) in a turn lane, and highway patrol with a vehicle pulled over (both parked) on the shoulder.  Interestingly, the moving truck was identified as a vehicle and displayed on the screen, though the parked vehicle(s) and the emergency vehicle incident was not.

A vehicle on the left exiting the highway is shown on the display ahead, because it is moving.

The vehicle on the left does not show up on the display because it is not moving.
Neither of these vehicles on the shoulder are displayed to the driver, including the officer standing outside of the vehicles.
It's possible (and highly likely) that the automation software did identify these vehicles, but filtered them out of the display.  Reviewing other Tesla blogs by users indicates this is a fairly well known issue; though it may be addressed in the update for version 9.  In any case I think it would be helpful if the vehicle at least identified them - and if it can highlight the emergency vehicle for the driver (and possible alarm) - that would certainly be a bonus.  Additionally, slowing down the vehicle since the officer was outside of his vehicle would have been welcomed behavior as well.  Because of the traffic configuration there was no opportunity to move over a lane to provide additional clearance for vehicles or the officer.

On the uphill climb the cruise control did pretty well, even recognizing that it needed to slow down turning into some of the curves, and in some cases accelerating coming out of the curves - which is precisely what an experienced driver on this road would likely do.

The one glitch that made me feel like I needed to temporarily disengage cruise control was when the vehicle accelerated coming around a curve to a hidden stop light.  The light has a flashing yellow light ahead of the turn so that drivers know to slow down when the light is flashing.  Normally mountain drivers coast at this point till they can see the traffic buildup and the status of the light.  Accelerating into the curve seemed hazardous enough that I tapped the brakes to turn the feature off - restarting it once we had passed the light.


Lane changing under Autopilot does not appear to be available on Highway 24 in Ute Pass (though it does under version 9).  Information in the operators manual indicates that there are various restrictions (poor visibility, sensor damage, vehicle in your blind spot, etc), one of which is "The road has sharp curves".  That may be the most likely reason - but it will need to be addressed to be viable for mountain driving in the future.  It's unclear if this will be considered level 5 autonomy or not.

Overall, with close management, the adaptive cruise control was primarily good for this type of mountain driving in optimal conditions (good visibility, clear roads, etc).  Autopilot though was not as good or reliable due to the algorithm's focus on keeping the vehicle centered in the lane, and in some cases edging uncomfortably close to the center line at inappropriate times.  The unavailability of the lane change function is also a factor in Ute pass, but not the primary concern - Autopilot needs to recognize when it should hedge more toward the center line, or the outside lane line to provide more space between vehicles or in some cases more clearance for roads with steep drop-offs, or to avoid debris.  Both features could embrace some idle slowing on blind curves where drivers tend to take their foot off the pedal and let it hover between the brake and the accelerator till what's around the corner becomes clear. V2V (vehicle to vehicle) communication may help with this concern in the future as well, but it will be several years before that additional information will be regularly available.



Additionally, it will be a welcome set of features showing progress toward self-driving autonomy when the vehicle recognizes speed signs and flashing warning lights as well as typical traffic lights parroting them on the driver's screen. This feature will initially help to potentially prevent drivers from inadvertently running lights, but will be a major step forward in providing drivers with confidence in the advancing technology.  I would expect this level of automation to be available for the level 4 deployment.  Some of the additional mountain driving requirements (such as hedging on narrow roads - especially in snow) may well end up being available only when level 5 autonomy is achieved.

So, it's pretty clear at this point that we're fairly far from reliable self-driving capabilities - and that may be the reason behind Tesla recently removing the 'Full Self-Driving Capability' package from its options due to 'confusion'.  This is more of a future upgrade of software and hardware to steadily push the technology closer to level 4 and level 5 over time; not a short term feature.  Fans of science fiction might feel like these features are very close to being a reality - when in actuality they are still several years off, with significant amounts of training and development to come.



That doesn't mean we should be disappointed though - this technology significantly improves driving safety now.  Properly used it will help guard driver's actions while protecting many lives today as well as tomorrow - as long as the driver stays focused and diligent.  We must keep patiently waiting for the technology to merge with our expectations.

One of the ways we can help to do that, is to drive and use the technology today; so that it will be ready for everyone in the future.

Next, will be the review of driving the vehicle on Pike's Peak Highway.

Tuesday, October 16, 2018

Tesla Deploys Massive New Autopilot Neural Net in v9, Impressive New Capabilities, Report Says

As reported by Electrek: While Tesla has held back its main new Autopilot feature with the release of version 9, it has still deployed a new neural net for Autopilot and according to a new analysis, it is a massively bigger neural net with impressive new capabilities.


Based on the new capabilities of Autopilot under version 9, we already knew that the new computer vision neural net had to be significantly updated.

It can now track vehicles and other objects all around the car – meaning that it makes better use of the 8 cameras around the car and not just the front-facing ones.

Now we have a better understanding of just how significant Tesla’s neural net update in version 9 is as TMC member Jimmy_d, a deep learning expert who has access to the software and has been releasing his thoughts on each update, has produced an interesting analysis on version 9.

Jimmy confirmed that Tesla has now deployed a new unified camera network that handles all 8 cameras.

He also listed a few other main changes:
  • Same weight file being used for all cameras (this has pretty interesting implications and previously V8 main/narrow seems to have had separate weights for each camera)
  • Processed resolution of 3 front cameras and back camera: 1280×960 (full camera resolution)
  • Processed resolution of pillar and repeater cameras: 640×480 (1/2×1/2 of camera’s true resolution)
  • all cameras: 3 color channels, 2 frames (2 frames also has very interesting implications) (was 640×416, 2 color channels, 1 frame, only main and narrow in V8)
Those changes add up to a much larger neural network that requires a lot more processing power.

Jimmy estimates that it might already be pushing the limits of the onboard computer – hence why Tesla is working on a computer upgrade.

He tried to communicate just how much bigger the neural net on v9 is compared to v8:
“This V9 network is a monster, and that’s not the half of it. When you increase the number of parameters (weights) in an NN by a factor of 5 you don’t just get 5 times the capacity and need 5 times as much training data. In terms of expressive capacity increase it’s more akin to a number with 5 times as many digits. So if V8’s expressive capacity was 10, V9’s capacity is more like 100,000. It’s a mind boggling expansion of raw capacity. And likewise the amount of training data doesn’t go up by a mere 5x. It probably takes at least thousands and perhaps millions of times more data to fully utilize a network that has 5x as many parameters.
This network is far larger than any vision NN I’ve seen publicly disclosed and I’m just reeling at the thought of how much data it must take to train it. I sat on this estimate for a long time because I thought that I must have made a mistake. But going over it again and again I find that it’s not my calculations that were off, it’s my expectations that were off.”
Based on his analysis, version 9 appears to be more than an incremental step change when it comes to computer vision.
Update: Elon Musk commented saying that Jimmy’s analysis is overestimating the increase capability which is estimate at closer to 400%:

Tesla deploys massive new Autopilot neural net in v9, impressive new capabilities, report says https://electrek.co/2018/10/15/tesla-new-autopilot-neural-net-v9/  by @fredericlambert pic.twitter.com/fBL9aDOudc





View image on Twitter
To be clear, actual NN improvement is significantly overestimated in this article. V9.0 vs V8.1 is more like a ~400% increase in useful ops/sec due to enabling integrated GPU & better use of discrete GPU.
The deep learning expert sees Tesla playing into its strengths with the update:
“Scaling computational power, training data, and industrial resources plays to Tesla’s strengths and involves less uncertainty than potentially more powerful but less mature techniques. At the same time Tesla is doubling down on their ‘vision first / all neural networks’ approach and, as far as I can tell, it seems to be going well.”
We are also starting to get a better understanding of what Autopilot can see through an effort from Tesla hackers.
Here’s a look at Tesla’s previous Autopilot software’s recognition of roadside structures:


ELECTREK’S TAKE
It’s really encouraging to see these important back-end upgrades to the Autopilot.

Tesla is really doubling down on the vision-based autonomous driving and making real progress on that front.

In terms of the actual experience on Autopilot, I have logged in about 100 km on Autopilot with v9 over the past few days and I’m truly impressed by the performance.

At first, I felt like it was ping-ponging more between the lines, something that never happened in the previous version, but it appears to have ironed itself out after 50 km or so.

Now the driving experience on Autopilot on the highway is virtually flawless for me.

The new neural net on V9 is also enabling some new renders on the center display (Model 3) and instrument cluster (Model S and Model X).

Tesla is now tracking vehicles all around the car and noting the difference between cars, SUVs, trucks, motorcycles, and it even renders pedestrians.

It certainly needs some improvements because I’m often getting those weird renders at low speeds or when stopped:






Tesla’s Autopilot on v9 shows great improvements when it comes to rendering the surroundings on the screen, but I’m often getting those weird bugs when stopped or at low speeds.

But those issues don’t affect the driving experience. I assume it will also improve with more data training the new neural net.

Of course, it’s always important to keep your hands on the steering wheel and stay attentive when driving on Autopilot regardless of these improvements.

Tuesday, October 2, 2018

Quintero One: This Hyperloop Pod Could Carry Passengers Next Year



As reported by Engadget: Hyperloop Transportation Technologies has unveiled its vision for a passenger capsule that will ferry people around the globe. The company has pulled the dust sheets from its first pod, called Quintero One, at Puerto de Santa Maria, the home of Spanish manufacturing partner Carbures/Airtificial.

Now that the vehicle has been shown to the world, it'll be sent to HTT's European HQ in France for "additional assembly and integration." The company expects that Quintero One will be pressed into service in a fully working commercial system at some point in the future.

We won't know much about how the pod will work, but some of the specifications the company has revealed shine a light onto its thinking. For instance, Quintero One measures 32 meters (105 feet) long, with the passenger cabin measuring 15 meters (50 feet).

We can assume that the propulsion, life support and other systems are buried both in the nose cone and behind the smaller passenger compartment. We don't know how many people HTT expects to seat in every pod, but its dimensions roughly correspond to a New York Subway car.

A 15.6-meter R188 Car has a maximum capacity of 188, which includes a significant number of standing passengers. Some back-of-the-envelope math suggests that a fully-seated pod could comfortably carry around 80 people. The company says that it's worth thinking about the pod as an "airplane without wings," which makes plenty of sense given the aerospace engineering necessary to make it work.

Both HTT and Airtificial (a partnership between aerospace companies Carbures and Inypsa) say that the pod is a landmark piece of engineering. Airtificial co-founder Rafael Contreras added that, because the pod uses an unprecedented amount of carbon fiber and composites its in its construction, it's the "safest transportation vehicle in the world."

If you're wondering, Quintero is a Spanish surname that means the "fifth," and is associated with farming as well as being the name of a cigar brand. It might just be an act of whimsy by the pod's engineers, or a subtle nod to the fact that Elon Musk called Hyperloop a "fifth" mode of transportation way back when.

HTT believes that the capsule will be ready to convey passengers before the end of 2019, and committed to building a test track at its French facility. The company has also built an insurance and certification framework that, it claims, will enable it to roll out its technology across the globe without interference or worry.

Monday, October 1, 2018

Mountain Driving the Tesla Model 3 - The Initial Driving Experience

Our new Tesla Model 3 outside of the iTRAK Corporate office.
I received my Tesla Model 3 last week, and have already started putting it through its paces here in the Pikes Peak region of Colorado.

For those that don't know me, my background is in electrical engineering, software development and systems design - specifically for end-to-end systems for vehicle tracking; something I've been doing for more than 20 years and have multiple patents and prior artwork in the field. Prior to that I worked on vehicle based terminals, GPS and RF communication systems for public safety agencies, as well as E911 communication and console systems employing early multiprocessor and redundant systems architectures.  I'm proficient in multiple computer languages and I'm an ongoing student in areas such as AI systems and Python.

The Pike's Peak region and surrounding area is an optimal area for testing, as it has some challenging traffic and terrain in areas like Ute Pass (Highway 24), standard city highway traffic along I-25, as well as a wide range of seasonal conditions such as heavy snowfall, heavy lightning storms, hard rains with significant run-off, occasional mixed icy conditions, high contrast lighting changes in canyon areas where GPS and wireless signals can be difficult to maintain, large temperature changes over relatively short periods of time as well as the hairpin turns at high angle and steep traffic at altitude such as can be found regularly on the local Pike's Peak highway - a specific area where automotive and brake manufacturers test new products.  More recently Pike's Peak (which locally is known simply as 'The Peak') highway has been featuring EV's racing through the challenging terrain - and just this year VW's EV broke the overall racing record.

I've been very interested in the evolution of electric vehicle technology, along with the advent of self driving systems for several years now.  Given their cost and the inherent risks involved in the technology, being on the bleeding edge of such development didn't seem like a good idea.  Instead I've waited patiently till I could see the technology entering the 'slope of enlightenment' period on the curve around the third-generation shown in the graph below:

In my opinion, the Model 3 vehicles from Tesla are at this stage now, though the self driving features are more likely between the 'peak of inflated expectations' and the 'trough of disillusionment'.  The vehicle that we purchased is configured for all wheel drive (a must for mountain driving) and the extended battery, along with the enhanced autopilot and the future self-driving capability (when it becomes available).  The range with the extended battery should be over 300 miles on a full charge which should allow reasonable access to the ski resorts in the winter (which has some fairly challenging back-country driving at times), as well as for longer drives along I-70 and I-25 to other States using the Supercharger network as needed.  I used the included 120V charging system for the past week till my home charging system could be installed by an electrician.

Like many, I initially ordered my Model 3 shortly after the ordering process started in March of 2016 without ever test driving one.  I've been keeping a close eye on news and reviews regarding Tesla's developing manufacturing process and release dates for the various versions and features - knowing I'd need a 4 wheel drive or all-wheel drive vehicle for my particular locale - and betting that a self driving firmware release was not far away.

In early September, I configured the vehicle and provided all of the information needed in order to procure a vehicle.  I expected a relatively long delay of 2-4 months for delivery, but was surprisingly pleased by a delivery date scheduled in late September.  There was a minor one week delay due to the break-down of the vehicle carrier in Wyoming, but that was relatively quickly resolved and easily scheduled around.

The paperwork at the dealership in Littleton was relatively quick, about 10-15 minutes before we were escorted to our vehicle.  The dealership was fairly busy, and they indicated they'd be trying to deliver almost 90 vehicles that particular day.  I had read some of the manual, and watched several of the videos, but that had been a couple of weeks earlier, so it was nice to review everything quickly.  After configuring our phones and a quick overview of the vehicle (don't use the manual opener for the doors, use the switch - use two hands to close the front trunk or 'frunk' and just lean on it to close the hood, orientation on the central touch screen of the vehicle, location of the tow bar, etc) we experienced some touch screen related issues.  This turned out to be related to the poor WiFi signal for the vehicle.  Once that was disabled (for the dealership's WiFi) it seemed to operate as expected.  After a visual inspection of the vehicle, I tentatively started the drive home.

The new vehicle after inspection in Littleton, CO.
My initial reaction was that the parking brake must be on - not because the vehicle wouldn't move, but because of how quickly it would start slowing down once my foot was off of the 'gas' pedal.  This impression is related to the regenerative braking system for the vehicle which instead of generating heat generates electricity to recharge the system - which as it turns out is an excellent adaptation for mountain driving for multiple reasons.  More on that later.

The autopilot and cruise control systems were not immediately available as they went through a calibration period while I initially drove the vehicle.  This was no problem as I really wanted a feel for driving the vehicle myself anyway.  The central map and navigation system (similar to the Waze app which I use when driving my other vehicles) was already setup with my home address so I was able to find out quickly how long it would take to get home in traffic, and how much battery I would likely have left or available to me at that time.  As the battery was not fully charged when I picked it up, and I have more than an hour of driving to do to get down to the Pikes Peak region - the 44% battery level on home arrival did prompt some 'range anxiety' as a first time driver, so I had my wife follow me closely as we drove home.  Rationally I knew this was not a problem, but my emotional intelligence on the matter needed a chance to catch up.  Building trust with the vehicle performance is a must.

I was able to get some music up and running using the streaming service at the stop lights while moving through Denver street traffic.  When 'The Killers' came on performing 'The Man', the lyrics "I got gas in the tank, I got money in the bank, I got news for you baby, you're looking at the man..." I thought - for the first time I'm driving without any actual 'gas in the tank'.  It had a bit of an eerie but futuristic synchronicity feel that I've run into with the vehicle several times afterwards.

Driving in street traffic was no different than driving any other vehicle, except for the learning curve on how to feather the 'gas' on the vehicle using the regenerative braking - sometimes referred to as 'one pedal driving'.  Once on the highway, the quickness of the vehicle was immediately evident - I bounced my head off the head-rest more than once merging quickly into highway traffic.

Related image
Tesla vision and sensor overview.
While in I-25 traffic heading home the calibration completed on the Autopilot and I started by using the cruise control feature.  The most noticeable difference from my prior experience using various prior vehicle's systems was the Tesla's ability to use its sensors and forward radar systems to keep a controlled distance from the vehicle just ahead of me (while visually showing vehicles around me in a kind of 'ghost' image on the center console) - matching its speed when lower than my max setting speed, and speeding up to the max speed when the vehicle exited the lane.  It did have one incident of an advanced warning on hard braking that occurred in the vehicle in front of the one I was following - not uncommon in the heavy traffic at that time of day.  It not only braked early itself enough to stop me from stomping on the brakes, but it also helped to keep driving stress of the surprise to a minimum.  Out of habit, I gravely remind myself that Seymour Cray (of Cray Supercomputer) was killed on this particular stretch of highway in 1996 in a car crash.  It's a serious reminder that automotive accidents make no exceptions for genius and we're all taking a significant risk as we regularly travel about in our daily lives.  I'd like to think he'd approve of the technology being used to help prevent those kind of accidents now, and into the future.  In any case, I was able to become comfortable with its use on the highway back to Colorado Springs, but fairly quickly determined that I wasn't comfortable enough to use it when exiting I-25 to highway 24 (though there may be settings to help optimize the performance I experienced).  I was able to re-engage it again once I had merged back into traffic heading west.  I was also able to use the feature up through 'Ute pass' without too many issues, though there was a hint of what I'd see later in the curves.

One advantage of the more standard cruise control mode (vs autopilot) is that you maintain steering through what is sometimes difficult and crowded traffic in Ute pass where road construction and 'scaling' of the surrounding mountain cliffs can be common - but without the need to focus on the stop and go traffic itself, which seemed to help with stress relief during the commute - something my wife would likely appreciate since she travels the pass on an almost daily basis and complains regularly about the erratic behavior of visiting tourists.

The trip home proved to be safe, uneventful, and very informative.

Within the first week I was able to take the vehicle on a two hour trip through Wilkerson Pass, Hoosier Pass, Breckenridge and Frisco to Silverthorne's Supercharger, as well as to make an initial drive up and back down Pike's Peak highway.  It was important to get these completed quickly while the conditions were optimal, as there is snow in the coming forecast for early October and I wanted to set a base line of testing under the best conditions first.

As this is meant to be a multi-part series review I will plan to follow up on those and additional drives and feature tests in the coming days and weeks.