DIPP: A Diffusion-based Potential Planner for Synergistic Navigation and Mapping

Yiqing Zhang^*, Yiqing Zhang, Yi Yang, Mengyin Fu

Beijing Institute of Technology
ICRA 2026

Paper: Coming soon arXiv: Coming soon

The introduced video about DIPP.

Abstract

Object-Goal Navigation (ObjectNav) requires an embodied agent to search for and reach a target object category in previously unseen environments using only onboard egocen- tric observations, which is a fundamental capability for long- horizon autonomous robots. Current Object-Goal Navigation methods typically discard environmental knowledge after each episode, limiting their ability to operate autonomously over long horizons. To overcome this limitation, we introduce DIPP, a diffusion-based potential planner that unifies navigation and mapping. DIPP generates two complementary potential fields: a navigation potential that directs the agent toward the target and a topological potential that captures the environment’s structural skeleton. The topological potential serves a dual purpose: it acts as an implicit structural prior for waypoint selection when fused directly with the navigation potential and, more importantly, enables the incremental construction of a persistent, explicit topological graph. This graph enables a hierarchical policy to select strategic, long-horizon waypoints, elevating planning from a tactical search to a strategic decision. We evaluate DIPP in the Habitat simulator on the Gibson dataset. Results show that DIPP achieves strong performance on standard ObjectNav metrics (SR, SPL) while constructing structurally accurate maps, evidenced by a high Node Recall score. Furthermore, leveraging the explicit persistent graph for hierarchical planning significantly boosts navigation perfor- mance. These findings demonstrate the effectiveness of DIPP in enabling embodied agents to build and exploit persistent spatial knowledge for long-term operation in unseen environments.

An example of DIPP, our framework for ObjectNav framed as a dual objective. Our core diffusion model, Topo- Diff, synergizes navigation and mapping by simultaneously generating a structural topological graph (blue) and a goal- oriented action plan (realized as the yellow path). This synergy enables the agent to build an abstract map while efficiently navigating from the start to the goal (red star).

Overview of DIPP. Left (online inference; repeated each replanning step): mapping updates an allocentric semantic map from observation and pose, then Topo-Diff generates a navigation potential and a topological potential conditioned on the map and target category (with classifier-free guidance using a null-goal branch). The topological potential updates the persistent graph, and the planner selects a long-horizon waypoint that the local motion planner executes into action. Right (offline supervised training): a two-stage curriculum pre-trains on topology-only targets to learn a structural prior, then fine-tunes jointly on navigation and topology targets.

Overview of the ground truth data generation pipeline. (1) A topological graph is extracted from the complete, ground-truth semantic map. (2) Object and topological poten- tial fields are calculated based on geodesic distances to goals and topological nodes, respectively. (3) These components are used to formulate training data pairs for the diffusion model.

Analysis of Denoising Steps at Inference. The left panel visualizes the generated topological potential field at different denoising steps, conditioned on the partial map (top-left inset). The potential field becomes progressively more defined as the step count increases. The right panel quantitatively shows the trade-off between computational cost (Time cost, in s) and mapping quality (Node Recall). A significant reduction in computation time can be achieved with only a minor drop in performance by reducing steps from 700 to 500.

A qualitative visualization of navigation with DIPP, tasked with finding a ‘toilet‘. The main images show the agent’s first-person view, with top-right insets displaying the local map and agent trajectory. The bottom row visualizes our model’s core outputs: the generated Topological Potential Field (Uˆtopo) and the resulting topological graph. (Steps 1-52) During initial exploration, our model generates a dynamic Topological Potential Field (Uˆtopo, green heatmaps), whose peaks directly guide the placement and adjustment of new topological nodes (yellow dots). (Steps 52-96) The value of this synergy becomes evident at the hallway junction (step 52), where the agent leverages the accumulated topological graph—itself a product of the potential field—to make an informed decision. This demonstrates our framework’s core feedback loop, where intelligent map construction provides the global awareness for efficient navigation to the goal.

Poster

BibTeX

@inproceedings{yiqing2026dipp,
  author    = {Yiqing Z, Tao W, Miaoxin P, et al.},
  title     = {DIPP: A Diffusion-based Potential Planner for Synergistic Navigation and Mapping},
  booktitle = {2026 IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
  publisher = {IEEE},
}