As of current release, Openpilot doesn't adjust for road weather (water, snow, mud) - an exception is that if you trigger traction control, it's supposed to disengage.
Openpilot won't make any changes to the model based on sensors on the vehicle related to the environment. (For instance if the rain sensor is triggering the windshield wipers or the external temperature is under freezing, Openpilot has no knowledge and doesn't take it into account.)
It wasn't trained to see standing water, does not avoid pot holes, manhole covers, or speed bumps.
It currently isn't trained to see traffic cones or traffic sticks, cannot see traffic lights, and cannot read any road signs (including but not limited to speed limit signs, school zone, yield, stop signs, construction zone, etc.)
One idea in the Openpilot community that has seen some discussion is that Comma wanted to design the system so that it can drive on any road without having any knowledge of the road. This makes perfect sense in that a human can drive in new environments safely without having knowledge of those environments. In practice this really only translates to a model that can follow lane lines and stop before rear ending the vehicle in front of you. There are some benefits, it nags you less than the factory system, and can steer with no lane lines as long as it is following a lead vehicle.
Designing a model that can always function in new environments also means they've made a decision against using GPS maps, and also the AI model doesn't learn previously driven roads in the way a human does. How the AI model adjusts to a road that has been driven repeatedly isn't really the same as, for instance, a human learning the curves of a road and preemptively adjusting the vehicle speed or lane position.
I feel that if a human can have existing knowledge of an environment, the Openpilot AI model should be able to use as much information as possible to make driving safer. (Including using live GPS maps or saved GPS data from previous drives.) Unfortunately this is not a goal of the project.
One of the biggest design issues, in my opinions, is that the weight of the physical components is suspended from the windshield. On a hot day the adhesive CAN come off. This happens on the older model more than the new one as the new one has a larger surface area for adhesive. On the Comma EON, in early May, I had come back to the vehicle to find the device sitting on my dashboard. I'm thankful that it didn't fall off when the vehicle was engaged as it could have steered me into a wall, barrier, or other vehicle.
If the physical part of the system were better designed, I may have less concerns about the system. For instance having a light weight camera on the windshield and having the processing unit somewhere protected from heat, that's a good start.
The software gripes, I think could be changed but at this point any large change to the core software is like trying to steer a large ship. If your idea is in conflict with Comma's idea, you have the option to fork the software, but your idea will never make it into the main community version.
Sent from my Pixel 4 XL using Tapatalk