Can Autonomous Vehicles Identify, Recover From, and Adapt to Distribution Shifts?
Authors: Angelos Filos, Panagiotis Tigkas, Rowan Mcallister, Nicholas Rhinehart, Sergey Levine, Yarin Gal
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We designed our experiments to answer the following questions: Q1. Can autonomous driving, imitation-learning, epistemic-uncertainty unaware methods detect distribution shifts? Q2. How robust are these methods under distribution shifts, i.e., can they recover? Q3. Does RIP s epistemic uncertainty quantification enable identification of novel scenes? Q4. Does RIP s explicit mechanism for recovery from distribution shifts lead to improved performance? To that end, we conduct experiments both on real data, in Section 4.1, and on simulated scenarios, in Section 4.2, comparing our method (RIP) against current state-of-the-art driving methods. |
| Researcher Affiliation | Academia | 1University of Oxford 2University of California, Berkeley. Correspondence to: Angelos Filos <angelos.filos@cs.ox.ac.uk>, Panagiotis Tigas <ptigas@robots.ox.ac.uk>. |
| Pseudocode | Yes | Algorithm 1: Adaptive Robust Imitative Planning |
| Open Source Code | Yes | Code and videos are made available at sites.google.com/view/av-detect-recover-adapt |
| Open Datasets | Yes | We first compare our robust planning objectives (cf. Eqn. (5 6)) against existing state-of-the-art imitation learning methods in a prediction task (Phan-Minh et al., 2019), based on nu Scenes (Caesar et al., 2019), the public, real-world, large-scale dataset for autonomous driving. |
| Dataset Splits | Yes | We use the provided train-val-test splits from (Phan-Minh et al., 2019), for towns Boston and Singapore. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | For all methods we use N = 50 trajectories, and in case of both DIM and RIP, we only optimise the imitation prior (cf. Eqn. 4), since goals are not provided, running N planning procedures with different random initializations. (...) We perform 10 trials per CARNOVEL task with randomised initial simulator state and the results are reported on Table 3 and Appendix B. |