Driver-WM: A Driver-Centric Traffic-Conditioned Latent World Model for In-Cabin Dynamics Rollout
Driver-WM: A Driver-Centric Traffic-Conditioned Latent World Model for In-Cabin Dynamics Rollout
Authors: Haozhuang Chi, Daosheng Qiu, Hao Su, Haochen Liu, Zirui Li, Haoruo Zhang, Chen Lv Date: 2026-05-06 Paper ID: openalex:2605.05092
Summary
Driver-WM is a latent world model that forecasts driver-centric in-cabin dynamics by causally conditioning on external traffic conditions. The model employs a dual-stream architecture that processes external traffic and driver states independently, using a gated causal injection mechanism to maintain temporal dependencies. This approach enables both long-horizon geometric forecasting of driver maneuvers and semantic recognition of behavioral states, facilitating controlled test-time interventions for safer L2/L3 automation.
Key Contributions
- Introduces Driver-WM, a latent world model specifically designed for forecasting in-cabin human dynamics conditioned on external traffic context.
- Implements a dual-stream architecture with gated causal injection to enforce temporal causality between external environmental and internal driver latent representations.
- Demonstrates superior long-horizon geometric forecasting for high-motion maneuvers and improved semantic alignment on a multi-task assistive driving benchmark.
Open Questions & Future Work
Key Concepts
- gated-causal-injection-mechanism: A mechanism that uses learned vector gates to modulate external contextual perturbations in a latent world model while enforcing temporal causality.
Archivist Review
I approved the gated causal injection mechanism as a reusable structural concept for directional conditioning in latent world models. I also approved a research question focused on the bottleneck of human-centered latent dynamics forecasting in autonomous systems. No datasets were approved as none were clearly named or unique enough for standalone note utility.
Approved Concepts
- Gated Causal Injection Mechanism: The core contribution for enabling directional coupling between external traffic and internal driver state streams while maintaining causality.
Approved Open Questions
- Anticipating Reactive Driver Dynamics: This represents a fundamental bottleneck in achieving safe, human-centered L2/L3 automation, as existing systems lack the capability to anticipate driver responses to specific environment triggers, leading to limitations in maneuver planning and safety-critical intervention.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.05092
- paper_id
- 2605.05092
- paper_source
- openalex
- domain
- nlp
- tags
- nlptime-series-forecastinglatent-space-modelingcausal-inference
- architectures
-
- datasets
-
- concept_slugs
- gated-causal-injection-mechanism
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-09T05:57:54Z