Abstract

The field of research in Autonomous Driving (AD) systems has witnessed an explosion of activity in recent years. Guided by the well-motivated separation of the overall pipeline into perception, prediction, and planning tasks, many works have demonstrated success in individually tackling challenging problems within the individual tasks. Especially in the field of prediction, recent works have achieved impressive results in forecasting the future motion of agents surrounding an autonomous vehicle, while producing competitive performance on large-scale public benchmarks. However, the promise of safer and smoother driving decisions with the help of performant prediction has not been delivered. Downstream planning components did not automatically benefit from these advances. The current state of research demonstrates that prediction and planning must not be seen as isolated tasks but instead as interdependent subtasks whose integration within the self-driving system is crucial to the overall safety and performance. This paradigm shift requires rethinking the design of the individual systems, the interfaces connecting them, as well as the individual and joint metrics and the benchmarks promoting overall system performance. For example, an important question is how to promote the bidirectional interaction between the two systems, where a planner is fed predictions, while a predictor can also benefit from the knowledge of future plans. This workshop is centered around these questions. Its goal is to bring together researchers from the diverged fields of prediction and planning to find a common ground for the design of a joint prediction and planning system. We want to engage in discussions on suitable interfaces between these two adjacent tasks and encourage future work to address both tasks as a tightly interconnected problem.

Call for Papers

We welcome paper submissions on novel and interesting approaches in the field of integrating prediction and planning. All papers will be carefully peer-reviewed based on their originality, relevance to the workshop topics, contributions, and technical clarity to ensure the high quality of the presented content. Accepted papers will be presented as posters. The poster session will start with a round of spotlight talks, where each author is given the opportunity to pitch their work in a 2-minute presentation. Every paper has to be covered in-person by one of the authors

Topics of Interest:

  • Planning-oriented trajectory prediction, including robust, reproducible, and (temporally)-consistent prediction methods.
  • Ego-conditioned prediction and world models
  • Interaction modelling between multiple road users in prediction and planning
  • Leveraging trajectory predictions in motion planning
  • Evaluating self-driving systems in highly interactive scenarios: datasets, metrics, and benchmarks
  • Uncertainties related to prediction outputs and their propagation to the motion planning task
  • Contingent motion planning based on anticipated future scenarios

Awards:

The program committee will honor the best paper among all accepted papers with a best paper award.

Important Dates:

We postponed our paper submisison deadline. There will be no further extensions!

  • Deadline for Paper submissions: August 1, 2024 September 7, 2024
  • Notification of acceptance: September 1, 2024 September 15, 2024
  • Deadline for Poster upload: September 15, 2024 October 7, 2024

Late-Breaking Work Paper Submission:

  • Deadline for Paper submissions: September 20, 2024
  • Notification of acceptance: September 27, 2024
  • Deadline for Poster upload: October 7, 2024

Submission Instructions:

  • Papers must not exceed 4 pages (excluding references and appendices)
  • Please refer to the main conference’s format guidelines and template for detailed instructions
  • Submissions are handled via CMT. You can submit your paper here

Tentative program

Time Speaker Topic
9:00 - 9:05 Organizers Welcome and Introduction
9:05 - 09:25
Zhiyu Huang, Nanyang Technological University Integrated Prediction and Planning for Interaction-aware Autonomous Vehicles Traditional autonomous driving systems often have separate prediction and planning modules for easier engineering, but this approach can reduce the effectiveness and interactivity of the planning module. Simply designing and improving the prediction module isn't beneficial unless it's related to the performance of the planning module. Therefore, integrating prediction and planning is key to achieving a more interactive autonomous system by tightly coupling the prediction and planning modules. In my talk, I will discuss several works on integrated prediction and planning (such as DIPP, DTPP, and HPP) to address this issue and develop more interactive autonomous driving systems.
09:25 - 09:45
Cunjun Yu, NU Singapore What truly matters in trajectory prediction Trajectory prediction is crucial for safety and smooth navigation in autonomous driving systems. However, there is a significant disparity between predictor accuracy on fixed datasets and real-world performance. We highlight the overlooked dynamics gap as a key factor in this disparity. In real scenarios, prediction algorithms influence the autonomous vehicle's behavior, which, in turn, affects other road agents, creating predictor-specific dynamics. This interaction leads to a gap between fixed dataset evaluations and actual driving conditions. We also examine the impact of factors beyond prediction accuracy, such as computational efficiency, and discuss potential solutions to mitigate this gap.
09:45 - 10:05
Peter Karkus, NVIDIA Interaction Aware Autonomous Driving with Differentiable Prediction and Planning Abstract: tbd
10:05 - 10:25
Ingrid Navarro, CMU Mining safety-relevant scenarios in real-world trajectory datasets to improve autonomous system robustness Scenarios sampled from real-world trajectory datasets are generally assumed to lack diverse and relevant situations essential for developing robust prediction and planning algorithms for autonomous systems. As a result, approaches for validating these systems have relied on on-road tests, which present potential dangers to drivers and vulnerable road users, as well as simulated experiments, which may not accurately capture real-world conditions and behaviors. While most scenarios and agent-to-agent interactions in trajectory datasets represent nominal behavior, we argue that safety should not only encompass situations in which agents act in a safety-critical manner but also situations in which safety-criticality and infractions are avoided through proactive actions. We believe mining for and leveraging these scenarios is an under-explored alternative that could improve prediction and planning stacks for autonomous systems. In our work, we have studied various strategies for characterizing safety relevance in scenarios drawn from real-world datasets as well as exploring “what-if” scenarios representing hypothetical situations in which proactive maneuvers were not performed. We have also proposed and explored various downstream applications to our strategies in autonomous driving and aviation. These include the creation of safety-informed distribution shifts and improving the robustness of trajectory prediction and navigation algorithms. The purpose of this talk is, thus, to discuss some of our findings as well as avenues for future research
10:25 - 10:30
Poster Spotlights Posters: TBA
10:30 - 11:00 Poster Session / Coffee Break
11:10 - 11:30
Alex Kendall, Wayve The Road to Embodied AI Abstract: tbd
11:30 - 11:50
Alexandre Alahi, EPFL 7 Foundational Principles for Embodied AI Abstract: tbd
11:50 - 12:00 Organizers Best paper Award & Conclusion