NextBestPath

NextBestPath

Efficient 3D Mapping of Unseen Environments

Shiyao Li1,2     Antoine Guédon1     Clémentin Boittiaux1     Shizhe Chen2     Vincent Lepetit1    

ICLR 2025

1École Nationale des Ponts et Chaussées     2Inria    
Paper Code Dataset and Toolkit

Left: Previous state of the art: MACARONS (see Related Projects).

Right: Our method.

NextBestPath enables the agent to perform efficient active mapping in unknown and complex indoor environments.

Overview

  • We introduce AiMDoom, the first benchmark to systematically evaluate active mapping in indoor scenes of different levels of difficulties.
  • We propose a novel next-best-path approach that jointly predicts long-term goals with optimal reconstruction coverage gains, and obstacle maps for trajectory planning.
  • Our approach achieved state-of-the-art results on both the AiMDoom and MP3D datasets.

AiMDoom Dataset

Random maps from the AiMDoom dataset, categorized into four levels by scene size and complexity, with 100 scenes per level, arranged from top to bottom: Simple, Normal, Hard, and Insane.

Indoor captures from AiMDoom.

Comparison between AiMDoom and prior indoor 3D datasets.

Architecture of NextBestPath

Overview of the next-best-path (NBP) framework. The model predicts a value map of coverage gain and an obstacle map, which are used for decision making to obtain a next-best path.

Training Procedure

Training procedure: We first gather training data from all training scenes using the current model, and then update the model with the new data. This process is repeated iteratively until the model achieves convergence.

Evaluation Results

Our NBP model can efficiently reconstruct unseen environments guided by predicted long-term goals, achieving state-of-the-art performance on both the AiMDoom (left) and MP3D (right) datasets

More Comparisons

Previous state of the art

Our method

Previous state of the art

Our method

Previous state of the art

Our method

Previous works mainly focus on using reinforcement learning (RL) or uncertainty prediction methods.

We improve the NBV-based SOTA method MACARONS with RL (the video in the second row), but it still underperforms our approach by a large margin.


BibTeX

@inproceedings{
      li2025nextbestpath,
      title={NextBestPath: Efficient 3D Mapping of Unseen Environments},
      author={Shiyao Li and Antoine Guedon and Cl{\'e}mentin Boittiaux and Shizhe Chen and Vincent Lepetit},
      booktitle={The Thirteenth International Conference on Learning Representations},
      year={2025},
      url={https://openreview.net/forum?id=7WaRh4gCXp}
      }

awesome webpage template