JPPD: Joint Prediction_Planning Diffusion with Differentiable Safety Guidance for Dynamic Obstacle Avoidance in Intelligent Transportation Systems
Researchers present JPPD, a joint prediction-planning diffusion framework that treats autonomous vehicle trajectory planning and pedestrian prediction as a single coupled problem rather than sequential steps. The approach uses differentiable safety guidance and conditional flow matching to improve safety metrics and runtime efficiency in shared-space transportation environments like sidewalks and pedestrian zones.
This research addresses a fundamental limitation in current autonomous navigation systems: the one-directional information flow between prediction and planning modules. Traditional architectures predict how pedestrians and other road users will move, then plan robot trajectories based on those predictions. This separation prevents the robot's intended behavior from influencing how the prediction model anticipates others will react—a critical gap in dynamic, interactive environments. JPPD resolves this by treating both prediction and planning as components of a single generative model using diffusion-based sampling with a causal Transformer architecture.
The technical contribution centers on differentiable safety potential guidance, which replaces heuristic post-processing with learned safety constraints that directly guide the sampling process. This enables the model to generate feasible trajectories without separate filtering steps. The use of conditional flow matching reduces computational overhead while maintaining trajectory diversity, addressing practical deployment concerns for resource-constrained platforms.
From an industry perspective, this work has significant implications for autonomous delivery systems, mobility-as-a-service platforms, and robotic applications operating in human-populated spaces. The evaluation framework—emphasizing near misses, blockage time, and induced participant deviation over traditional error metrics—reflects real-world operational requirements that matter to deployment stakeholders. The validation across simulation, naturalistic data, and actual ROS/Nvidia Orin hardware demonstrates maturity beyond theoretical contributions.
This approach could influence how autonomous systems researchers design safety-critical navigation stacks, shifting from modular decomposition toward end-to-end, interactive modeling. The work particularly benefits developers building low-speed autonomous platforms where safety certification and shared-space efficiency are paramount competitive factors.
- →Joint prediction-planning diffusion eliminates information bottlenecks by sampling robot and pedestrian trajectories from a coupled distribution rather than sequentially
- →Differentiable safety potential guidance replaces ad-hoc collision avoidance with learned constraints that directly guide trajectory generation
- →Evaluation metrics prioritize operational safety indicators like near misses and blockage time over traditional displacement-based error measures
- →Conditional flow matching reduces inference latency while preserving multimodal trajectory diversity for real-time deployment
- →Hardware validation on Nvidia Orin demonstrates practical feasibility for autonomous platforms operating in shared pedestrian spaces