Vendor MaxSavr® System for recording & privately replaying Autopilot/driver data

[Telanon]

Has anyone driving your Tesla experienced a strange or dangerous situation, or even a crash? (If you weren’t there, unless there was a crash or traffic ticket, how would you even know?)

If so, were you later able to get a clear understanding what the driver &/or Autopilot did in the critical seconds just before it?

We believe it’s currently too hard and sometimes impossible to know. That’s why to solve that problem we’re building MaxSavr® Driving Data Systems. We have a prototype working, and now need insights from potential customers.

So if you own a Tesla and this sounds interesting, please click here to see details and complete a short interest survey: MaxSavr® Driving Data Systems . It’s only 8 questions, nearly all only require a click or two to answer, and it shouldn’t take more than 10 minutes. After you finish and click Submit, you'll receive an e-mail with the questions and your answers.

Here's an example of MaxSavr's full capabilities.

Situation Replay details:
Tesla Model X with Autopilot 2.5 in control, approaching slower motorcycle at night

MaxSavr® Visualizer radar-view explained, including tailgating detection & Autopilot:
Tailgating detection lights at bottom left are based on distance and Tesla's speed
-- yellow is "close"; orange is "closer"; and red is "too close"
Parallel green lines are Tesla's current path straight ahead from 0 to 200 meters
Tesla is at position 0 and currently driving straight ahead
Forward-moving object detected straight ahead (motorcycle in video view)
-- is small blue circle when in Tesla's current path
-- becomes small blue square when it leaves Tesla's current path
Small grey squares are out-of-path stationary objects
Autopilot is active (when both Adaptive Cruise & Lane Keeping are "Yes")
-- driver expected to immediately take control in case of Autopilot error
-- limited to set speed (70 MPH here, and Accel Pedal % = 0 shows driver did not override)
-- reduces speed for slower vehicle in its path
-- speeds up when path is clear
-- keeps Tesla centered in lane

 

[Telanon]

I want to point out two more things not in the initial post.

First, completing the survey also lets us contact you when the right MaxSavr® product for you is available, based on your survey answers. So if you skipped the survey earlier but that makes a difference, please complete it now.

Second, the Situation Replay example showing Autopilot's response when approaching a slower vehicle ( getting "too close" in our opinion, shown by the red alert light) is a lesson to the vehicle owner (and driver if someone else). It's to watch carefully in such situations and be ready to take control if needed, especially if Autopilot following distance is set very low.

Otherwise, here's a vivid example of what can go seriously wrong: https://twitter.com/greentheonly/status/1360622235053133829 . It's a video created by green from a crashed Tesla's Autopilot computer. Following distance was set to only 1. After the close cut-in the driver didn't intercede quickly to provide more following distance, and then the driver also didn't immediately brake when needed before Autopilot did. (NOTE: green investigated the braking messages and later released this corrected version: https://twitter.com/greentheonly/status/1360723412612767748 .)

Too-close following is only one dangerous driving behavior that will trigger MaxSavr to record data for later review. Harsh braking, high acceleration or speeds, and actual crashes are others. It's all for private review by the Tesla owner as feedback for safer driving, and uses a secure chain of custody we can certify if needed for evidence.

 

[Telanon]

A one-question poll has been added at the top of this thread. If you own a Model S, X, 3, or Y and haven't completed our 10-minute survey, please consider answering the poll instead. The full survey gives us more details, but even just the results of the one-question poll will help. Thanks!

 

[Telanon]

Here's another example of MaxSavr's capabilities, this time with enlarged text.

Situation Replay details:
Tesla Model X with Autopilot 2.5 in control, following car towards roadblock

MaxSavr® Visualizer radar-view explained, including tailgating detection & Autopilot:
Parallel green lines are Tesla's current path straight ahead from 0 to 100 meters
Tesla is at position 0 and currently driving straight ahead
Forward-moving object detected straight ahead (red car in video view)
-- is small blue circle when in Tesla's current path
-- becomes small blue square when it leaves Tesla's current path
Oncoming vehicle in adjacent lane detected (light-colored car in video view)
-- is small red square (red indicates oncoming object, square means not in current path)
Small grey squares are out-of-path stationary objects
Autopilot is active (when both Adaptive Cruise & Lane Keeping are "Yes") until it stops vehicle
-- driver expected to immediately take control in case of Autopilot error
-- reduces speed for slower vehicle in its path
-- briefly speeds up when path seems clear, then applies brakes when roadblock is detected
-- keeps Tesla centered in lane

 

[Cheburashka]

It seems like it may be a good idea, but my sentiment is that for most accidents the dashcam footage is good enough.

 

[Telanon]

Thanks for your comments.

What dashcam alone doesn't reveal is what the driver did versus Autopilot/AEB. Even the driver might not be sure, especially later.

Also, if the Tesla owner isn't in the car and the driver doesn't mention a strange or dangerous event that won't get the owner's attention anyway (crash or traffic ticket), the owner won't know to look through the video, especially not when in the video. MaxSavr would automatically save the data including timestamp for crashes, near-crashes, and extreme tailgating (if a compatible installation was chosen), and any time the driver notices something unusual and taps the horn.

The last advantage is that secure chain of custody will be preserved for the data. (It gets encrypted before storage in the MaxSavr device in the car, and not decrypted until it reaches our server applications.) So it should be allowable as evidence if needed for legal purposes to understand driver versus Autopilot/AEB actions in a given situation. Video alone can literally show what happened, but not explain the driver versus automated actions happening at the same time.

Those general lessons are useful, but even more valuable is timely feedback for a particular driver from his or her own recent driving. The Tesla owner can tell from the timestamp when each event occurred, and use it to identify the driver if it was someone else.

 

[bedoig]

I may be misunderstanding after just skimming through, but what difference does it make to know if the actions were initiated by the driver or autopilot? All liability is going to rest with the driver for quite some time still I imagine.

 

[Telanon]

That's my understanding too, but I've seen multiple reports of Tesla owners or families attempting to get such information from Tesla after an accident, and only getting it after legal action if at all. I assume they want the data to see if it supports any reason for challenging that position.

Getting those insights are useful in non-crash cases, too. The idea is to get convenient insights into what happened in dangerous or strange situations, to better prepare drivers for anything similar in the future. The better they understand how Autopilot has reacted in the past, the more prepared they should be to anticipate what they should do in the future.

The MaxSavr tailgating video was an example. It alerts viewers that Autopilot by itself can get what we consider "too close" to a slower vehicle before it establishes an appropriate following distance. Seeing that example, a driver should be on the alert to similar situations. Green's video from a crashed Tesla shows what can happen in such situations when the driver isn't sufficiently aware s/he should be alert to take over from Autopilot then.

 

[bedoig]

It's an interesting project. I don't think I'd be interested other than just in passing for the fun of looking through the data here or there. Details on pricing (initial costs and any ongoing fees) would be useful, as would information on what hardware is involved, and whether any software modifications to the car are necessary.

 

[Telanon]

What I can say with certainty is there are no software changes needed to the car. Our approach simply reads and interprets data from the Tesla, whatever software it's running.

What I can add with near-certainty is about ongoing fees. We plan to use WiFi to transmit the data from MaxSavr in the Tesla to our server apps, whenever it is within range of the WiFi you normally use for Tesla software updates. So there's no cost to either you or us for transmitting the data. We do use resources to convert the data, so it won't be entirely cost-free for us, but we plan to offer some reasonable level of monthly activity at no monthly cost. For customers desiring higher levels of activity, we'll offer activity-based monthly plans.

That leaves your questions about hardware and prices for it and installation. The only hardware you need is the MaxSavr module we provide and the wiring assembly needed for your Tesla. Below are answers about pricing, both for the module and installation.

We're pre-launch now, with a prototype running and decisions to make on which features, Tesla models, and viewing approach has enough interest to support first, and at what price. The survey is the way interested Tesla owners can help us make those decisions. It explains all that, and uses a non-competitive product, Ingenext's Bonus Module that's already available for Teslas, as a reference for value, pricing, and installation alternatives.

Essentially, with the survey we're saying "You tell us about your Tesla, your mobile devices, the installation alternative(s) acceptable to you, and how much you think is an acceptable price. We'll decide what has enough support to launch, and get back in touch when your version is ready."

Even if you don't want to complete the survey, you can go through it to learn all available details, and simply not click "Submit" at the end. We won't get any of whatever answers you gave, but you'll at least learn all the details in the survey not yet duplicated in this thread. If you want to submit, great -- it will help us a lot. If you don't, but are then able to answer the poll question, that's helpful too, but not nearly as much.

As for the price for the module, we identify another product for Teslas and ask you to tell us how you feel MaxSavr compares in terms of value, and then what you feel would be an acceptable price.

As you see from a survey screen below, installation varies depending on your model, in some cases whether you're interested in tailgating detection, and whether you want professional installation or are willing to do self-installation knowing the implied time & effort based on the estimated professional cost. For installations without tailgating detection, it's possible to change the installation later to support it.

 

[Telanon]

One survey responder requested "more data and actual video of system for more than 15-20 seconds". Here's an example that includes something that puzzles many Tesla drivers -- phantom braking.

First, here's the Situation Replay version that we think is best for MaxSavr. It has the MaxSavr Visualizer view side-by-side with a video and synchronized. Notice you hear an alert just after Autopilot starts slowing the car, and it's a puzzle to explain. We think Autopilot visually projects the road just over the hill straight through where the power pole is located, and starts slowing plus alerts the driver. Then the car crests the hill and it's obvious that the road actually curves a little to the left, not straight to the pole, and Autopilot starts speeding up again.

Next, if you instead got the longer video and MaxSavr data, they look like this. You would need to match timestamps between the two, since they aren't shown together and synchronized, but this particular video doesn't show the timestamp so that would be difficult. We think a short synchronized version covering the necessary time to understand an important situation, that we call a Situation Replay, is the right tradeoff. Data files like these can get very large so we want to only capture and process what's needed each time there's a trigger that indicates an important situation. It's automatic for crashes, near-crashes, and too-close tailgating (if installation supports), and when the driver taps the horn for anything unusual like phantom braking.

 

[Telanon]

Here's another Situation Replay for phantom braking, this one recorded yesterday. (Ignore the timestamp in the bottom-right corner of the video; the dashcam's battery had died and timestamp wasn't reset before using it, so it thought it was New Year's Day in 2016.) In this case it wasn't a true "phantom" since it was clear what caused the braking. It was simply unnecessary braking. As the Tesla driver and the driver of the red car just ahead could tell, the turning vehicle wasn't a realistic collision risk.

Situation Replay details:
Tesla Model X with Autopilot 2.5 in control, following car & approaching another car turning in its direction

MaxSavr® Visualizer radar-view explained, including tailgating detection & Autopilot:
Tailgating detection lights at bottom left are based on distance and Tesla's speed
-- yellow is "close"; orange is "closer"; and red is "too close"
Parallel green lines are Tesla's current path straight ahead from 0 to 200 meters
Tesla is at position 0 and currently driving straight ahead
Forward-moving object detected straight ahead (red car in video view)
-- is small grey square at first before it turns in Tesla's direction, since it was a stationary object
-- becomes a red square at first when it begins moving (red indicates oncoming object, and square is outside current path)
-- briefly becomes a red circle when directly ahead (red still indicates oncoming, and circle means within current path)
-- even though it only crosses over Tesla's current path, radar considers it "oncoming" because it's moving in Tesla's direction
Other small grey squares are also stationary objects outside Tesla's current path
Autopilot is active (when both Adaptive Cruise & Lane Keeping are "Yes") at all times in this Situation Replay
-- driver expected to immediately take control in case of Autopilot error
-- applies brakes to reduce speed, due to judgement turning car briefly in current path is a collision risk
-- driver quickly presses accelerator pedal because he realizes the turning car is headed to a lane outside current path
-- driver pressing accelerator pedal "overrides" to cancel Autopilot braking and allows driver to control speed instead
-- keeps Tesla centered in lane