JPPD framework unifies prediction and planning with differentiable safety for robot navigation
New diffusion model lets robots jointly predict and plan paths avoiding obstacles safely.
Autonomous platforms navigating shared spaces — like sidewalks, hospitals, and airports — have traditionally split motion planning into two sequential phases: first predict where pedestrians, robots, and micromobility users will go, then plan a safe robot path. This one-way pipeline ignores that the robot's chosen action influences everyone else's future trajectories. A new paper from researchers Jiahao Wu and Shengwen Yu (arXiv, June 2026) proposes JPPD, a joint prediction-planning diffusion framework that treats both the robot's plan and all participant trajectories as a single conditional generation problem. At its core is a causal Transformer with cross-trajectory attention, which samples the entire multi-agent future from one coupled distribution — making the robot's plan directly affect predicted participant behaviors.
The framework introduces differentiable safety potential guidance, a time-varying occupancy-probability potential whose gradient steers the joint sampler away from collisions, replacing traditional heuristic repulsive post-processing. Conditional flow matching accelerates inference while maintaining multimodal trajectory diversity. Evaluation goes beyond standard displacement errors, focusing on real-world operational metrics: near misses, blockage time, induced participant deviation, hard-braking events, and embedded latency. Experiments in scenario-grounded simulation, naturalistic pedestrian replay, Isaac Sim validation, and ROS/Orin deployment show that JPPD significantly improves tail safety (rare but critical events) and runtime efficiency compared to the separated prediction-then-planning baseline.
- JPPD uses a causal Transformer with cross-trajectory attention to jointly sample robot and participant trajectories from one distribution, enabling bidirectional influence.
- Differentiable safety potential guidance replaces heuristic repulsion with a gradient-based method, improving collision avoidance in dynamic environments.
- Evaluated on tail safety metrics (near misses, blockage time, induced deviations) in simulation and real deployment, outperforming traditional pipelines.
Why It Matters
Enables safer autonomous navigation in crowded shared spaces by unifying prediction and planning with principled safety constraints.