-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
diplomat33
Average guy who loves autonomous vehicles
I think it does. DirtyTesla mentions it in his first impressions video of V9.
diplomat33
Average guy who loves autonomous vehicles
Chuck doing the same route on V9 for comparison.
scottf200
Well-Known Member
mikes_fsd
Banned
VS
Probably worth at least rereading what you wrote before posting.
To be clear, on top of the refinements, they've changed the underlying stack!
That's not a "minor" thing.
mikes_fsd
Banned
Elon's comment on observations made by Earl
Leading the autonomy software team for the Tesla Autopilot.
My team's main focus areas are:
- Creation of large scale automatic ground truth pipelines to train neural networks with massive amounts of diverse, high-quality data. Use this fleet-learning approach to replace potentially brittle run-time algorithms with robust learned models.
- Developing an accurate and detailed geometric and semantic understanding of the world using the best of both machine-learned and engineered models.
- Building robust, causal, predictive models for other agents in both geometry and semantic state spaces.
- Decision making, motion planning, and control modules using state-of-the-art AI techniques including methods for high-dimensional search, trajectory optimization, reinforcement learning, model-predictive control, etc.
Both a manager and a technical contributor, I've brought up Tesla's successive Autopilot HW2 and HW3 generations of computers and software stacks, integrated them across all Tesla vehicle types and pushed them all the way to mass-production. I've been regularly shipping software updates to hundreds of thousands of vehicles across the world, including compute optimizations, new features, and system stability fixes.
I've been scaling Tesla Autopilot's software- and hardware-in-the-loop continuous integration infrastructure, and developed productivity tools to accelerate our R&D team's development cycles towards full-autonomy.
I currently report directly to Tesla CEO.
In particular, I currently lead:
- Overall System Software & middleware (C/C++ middleware, IPC, process scheduling, Logging, Watchdog, ...)
- Computer Vision system software (GPU kernels for post-processing, Neural Network integration, C++ Compute Graph Framework for efficient compute scheduling across multiple devices)
- Camera software stack (across all Tesla vehicle types)
- Platform Software (Linux kernel/drivers, security, power, board bring-up)
- Continuous Integration infrastructure (automated & on-demand support of regression tests, performance tests, Simulation tests, with scheduling on either x86 emulation, as well as on true hardware-in-the-loop setups)
- Build System (including remote-caching)
- Performance & optimization (responsible for the Autopilot framerate across all platforms)
- Telemetry (on-vehicle data capture software, and back-end ingestion services)
- Machine Learning infrastructure (training stabilization & scalability, workflow automation)
- Tools (sensor clips visualization, data plotting for logs analysis or live debugging)
Both guys seems pretty busy. I guess developing the FSD software stack is a pretty massive task…
So basically:
Andrej: Train neural networks
Ashok: Generate dataset for neural networks
Milan: Software 1.0 surrounding software 2.0 to deploy on HW3 and DOJO
Last edited:
mikes_fsd
Banned
diplomat33
Average guy who loves autonomous vehicles
Looks like little or no change to reliance on map data, so I would expect similar continued failures that others have noticed especially in downtown areas. Notably, ending up in a left-turn only lane at 10:30 where OSM believes there's 3 lanes with no turn lane attributes as well as changing lanes during an intersection at 12:25 because map data says 4 lanes without indicating which direction has how many, so Autopilot probably assumed 1 and cut-off/"merged" in front of the other car.
Last edited:
Daniel in SD
(supervised)
Looks like it makes driving much more relaxing!
I do wonder whether they might disable some maneuvers for a public release. I suppose the issue with that is it might make people more complacent for other maneuvers since they might assume that the car will tell them when it can't handle something.
mikes_fsd
Banned
This is why it's in beta with a limited user base!
I know you don't like this, don't agree with their approach, but is it really that hard to actually remove the snark and focus on the functionality?
What's the path forward?
Does this mean the whole thing should be scraped?
Can they solve this within their system?
On the other hand, he said that the intervention frequency was about the same as before.
Starting from 16:26 fsd doesn’t see monorail pillars and is about to collide with them.
18:00 fsd doesn’t see planters and is about to collide with them.
I don’t see any huge improvement, just incremental.
Last edited:
Daniel in SD
(supervised)
Those are good questions. I guess I don't understand what the goal is. I assume they'll be able to do way better than they're doing now for that maneuver. I'm not sure I understand the point of real world testing is when I'm sure this also fails in simulation. I think they should scrap the idea of doing unprotected lefts like that until the safety is greater than a human. Monitoring the system while it's doing that maneuver looks very difficult. Obviously fearing for your life helps keep focus but what happens when it gets 100x better? Will people still have the fear of death or will they lose focus for a split second with disastrous consequences?
Success rate was 0/4. This was supposed to be the vast improvement. Smh.
As someone who is driving around 3,400 miles per month for work, I would love to see this technology ready but that is so far from reality right now
Daniel in SD
(supervised)
In Tesla's defense Elon only said it would be a "foundational" improvement which to me implies that the actual performance might not be improved yet.
Similar threads
