Overview

Situation cognition in driving is the active perception and prediction of the intentions and future behaviors of involving agents and facilitating the inference of spatiotemporal constraint for behavior decision-making, motion planning, and motion control of autonomous vehicles. In the workshop, we would like to discuss how recent advances in situation cognition from the AI communities could benefit driving autonomy research, e.g., incorporating foundation models, scene graphs, causal theory, etc., potentially facilitating fully autonomy of driving systems and enhancing the interpretability, safety, robustness, and generalizable ability of autonomous vehicles. 

Nowadays, the situation cognition computing in the open traffic world faces various environmental disturbances, causal confusion, and various physical properties and rules. Current paradigms fall short of the traffic situation reasoning that vehicles need to consider, such as human experience, real-world knowledge, situation causal relations, etc. These issues may be alleviated by integration with driving systems, where real-world driving experiences that contain rich sensory input and introspection can potentially bring driving autonomy to the next level. 

The workshop aims to create an inclusive and collaborative platform for professionals, researchers, and enthusiasts alike to exchange ideas, share experiences, and foster meaningful connections within the autonomous driving community. It will feature a mix of presentations, open panel discussions, and results and demos from our Visual-CognitiveAccident-Understanding (Cog-AU) competition challenge. Five invited speakers will share their related research, thoughts, and experiences at the intersection of autonomous systems, with broad coverage of topics such as available datasets, benchmarks, closedloop driving, foundation models, and more.

                                                             Call for Papers

Autonomous driving has the dual attributes of strategic technology and emerging industry and has long been highly concerned by local governments, industry, and academia. In recent years, continuous breakthroughs in learning technology have pushed autonomous vehicles into a new stage of development. In this way, existing autonomous driving systems still cannot omit manual supervision. Driven by the increasing scale of model parameters (e.g., LLMs, VLMs) and the massive expansion of data, deep learning technology promotes the continuous iterative upgrade of driving system performance along three technical routes: modular cascade, end-to-end perception-decision integrated model, and foundation model empowerment. However, bottlenecks such as interpretability, generalization, robustness, and causal confusion still exist and are increasingly obvious, seriously affecting the industrialization process of autonomous driving systems.

Traffic scenes are uncertain and open. The human driving process is a dynamic process that continuously recognizes the traffic situation and generates driving behaviors, which is obviously different from the current perception-to-control calculation process of autonomous driving. Humans learn, interpret, reason, and remember situations and corresponding driving operations from their interactions with real traffic environments, conduct inductive reasoning, summarize common sense rules, and accumulate driving experience. When humans drive, the process of triggering driving behavior by the current traffic situation is the result of cognitive mapping between human experience and the current situation.

How to transfer the scene semantics, abstract representations, and mapping mechanisms of driving experience formed by humans to autonomous driving systems, so that they can recognize and respond to highly dynamic and highly random traffic situations like humans? Obviously, the "representation+calculation" paradigm followed by current data-driven deep learning cannot adapt to the closed-loop characteristics of intelligent driving perception-motion, especially the unique introspection and reflection mechanism of human driving strategy learning. This requires exploring new types of autonomous driving systems with human-like cognitive mechanisms (abstract representation of situations and human-like policy generation for introspection and reflection) from the perspective of computational thinking. 

The workshop aims to create an inclusive and collaborative platform for professionals, researchers, and enthusiasts alike to exchange ideas, share experiences, and foster meaningful connections within the human-like autonomous driving field. It will feature a mix of presentations, open panel discussions, and results and demos from our VisualCognitive-Accident-Understanding (Cog-AU) competition challenge to facilitate the development of safety-critical driving scene understanding. In this challenge, we plan to host the challenge over the span of a few months before IROS 2025 and present the first stage of results at the workshop. We will release a novel human-gaze-aligned egocentric accident video benchmark collected in driving scenes. Specifically, the benchmark contains well-annotated human-gaze maps, object bounding boxes, accident videos, text descriptions of accident reasons, prevention advice, and accident types. The challenge will be open to anyone interested in the field to participate and we will summarize the results and invite the winning teams to present their work at the workshop. More information on the Cog-AU Challenge can be found on the workshop website.

Five invited speakers will discuss their related research, thoughts, and experiences in various aspects at the intersection of AI and autonomous driving systems, with broad coverage of topics such as available datasets, benchmarks, software stacks, world models, spatial reasoning, foundation models, and more. Additionally, this workshop calls for workshop papers to absorb cutting-edge works with poster presentations. 

Topics of interest span a diverse range across the field of autonomous driving and other disciplines and include but are not limited to the following:

• -embodied driving intelligence

• -spatial and causal reasoning

• -social concept reasoning in driving scenes

• -human-vehicle collaborations

• -memory modeling and retrospection in driving autonomy

• -foundation models for autonomous driving

• -vision-language-driving world models

• -Real-to-sim-to-real gap in fast-driving testing

Invited Speakers

Organizers

Paper Submission

Papers to IEEE/RSJ IROS 2025 can be submitted on the  [OpenReview Portal]. The page limit is 6 pages, with up to 2 extra pages (with extra page charge).  The page limit includes the references, appendixes etc.  

All papers must be submitted in PDF (up to 6MB) and must follow the IROS double column format. Information and templates are available [here].