YassineYousfi

We are aware of this issue, and working to fix it

The DCTR (and JRM) feature extractor is used in `abba/rich_models/` we extract the DCTR features for the entire dataset and train an FLD ensemble for each quality factor. We use...

Just happened in CI again in 3799fe4.

why is the `(v + velocities[0]) / 2)` needed?

That makes more sense. Or you can integrate the speeds up to t=delay using np.cumsum if you want to quickly experiment without changing the messages.

We are looking into this. Some turn cutting comes from higher in the planner stack which we are also looking into.

The current supported onnxruntime version is 1.12.1 https://github.com/commaai/openpilot/blob/master/poetry.lock#L2557 We currently only support that (1.13.1 might also work) Why do you need onnxruntime==1.14.0? We can update the models to work with...

@sshane this works much better than master on the bronco, lgtm