Planning with an Ensemble of World Models

Arun Balajee Vasudevan, Neehar Peri, Deva Ramanan

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Motion Planning, Evaluating Motion Planning, World Models

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This work propose a set of world models for better motion planning and evaluation of planners.

Abstract: Motion planning is of critical importance for safe navigation in complex urban environments. Historically, motion planners (MPs) have been evaluated using procedurally-generated simulators like CARLA. However, such synthetic benchmarks are not reflective of real-world multi-agent interactions. nuPlan, a recently released MP benchmark, addresses this limitation by augmenting real-world driving logs with closed-loop simulation logic, effectively turning the fixed dataset into a reactive “gym” simulator. We evaluate the quality of nuPlan’s Default-Gym and find that it does not accurately reflect real-world human behavior, particularly for cities with unique driving behaviors (e.g., Boston drivers tend to be more aggressive than Pittsburgh drivers). Therefore, we propose city-specific gyms (e.g., a Boston-Gym and Pittsburgh-Gym) to evaluate planning performance. Evaluating a state-of-the-art planner with our proposed ensemble of gyms yields a drop in performance, suggesting that a good planner must adapt to different environments. Leveraging this insight, we present City-Driver, a model-predictive control (MPC) based planner that unrolls a city-specific world model that adapts to different driving conditions. Our extensive experiments demonstrate that City-Driver achieves state-of-the-art results on the nuPlan benchmark, reducing test error from 6.4% to 4.8%.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 369