It is crucial to quantify the uncertainty of these dynamics models when using them for planning.

We propose a method called clustered conformal prediction that clusters together classes having "similar" conformal

Previous work has focused primarily on developing ML-based "meta-learners" that can provide point estimates of the conditional

Full (exact) conformal set vs. split or cross-validated conformal set Non-connectedness of the conformal prediction set. This was initially suggested in [18, Remark 1]. We follow the actual practice in the literature [14, Remark 5]. We did not observe violations. We will also summarize the proposed algorithm in a direct pseudo-code.

Our theory both simplifies and strengthens existing analyses in online confor-mal prediction.

Backdoor attack is a common threat to deep neural networks.