FAME: Multimodal Robot Exploration via Conditional Flow Matching for Active 3D Mapping

From Single-Target Prediction to Multimodal Exploration

Active 3D mapping requires a robot to reconstruct the 3D structure of an environment from observations and actively decide where to move next to gather information, aiming to rebuild the scene as completely as possible with minimal motion and sensing cost. However, the robot can only observe a part of the environment at any time. Even when two scenes share similar local structures, the unexplored regions may differ completely. If a model predicts only a single long-term exploration target based on past experience, it can easily choose the wrong direction in unfamiliar environments, missing other valuable areas.

The research team points out that under partial observability, long-term exploration targets often exhibit multimodal characteristics : for the same current state, multiple reasonable future exploration directions may exist simultaneously. From a theoretical perspective, when several plausible targets lie in different directions, a single-target model trained with squared error tends to output the average position of those targets—a location that may not be a valid exploration goal and could even fall inside an obstacle. Therefore, instead of forcing the model to predict one target, FAME lets it consider multiple possible exploration directions at once.

Generating Multiple Exploration Hypotheses with Conditional Flow Matching

To enable the robot to generate multiple possible exploration targets simultaneously, the team introduces Conditional Flow Matching (CFM) . Given the map the robot has built so far, the model starts from random noise and progressively generates several distinct target positions and orientations, called Anchors . Each anchor represents a potential long-term exploration direction—for example, one anchor may lead into a left-side room, another may continue down a corridor, and another may point toward an uncovered local area.

Unlike methods that directly generate full exploration trajectories, FAME only predicts the target position and orientation . The model proposes a few candidate destinations, and a path planner then computes how to reach each one given the current obstacle map. This design preserves multiple exploration directions without requiring the model to output complete paths.

Not Just More Paths—Learning How to Choose

Generating multiple anchors is not enough; the robot must also judge which paths are feasible and which one to execute. FAME employs an A*-based path planner to generate a candidate path for each anchor using the current obstacle map. The system then clusters similar paths based on path endpoints, geometric shape, initial execution behavior, and obstacle-avoidance topology. This groups anchors that lead to similar trajectories into the same exploration mode , reducing redundancy in later decision-making.

Finally, FAME performs hierarchical selection : first at the exploration-mode level, then within the chosen mode at the individual-path level. The first layer weighs potential exploration gain, risk, uncertainty, cost, and diversity; the second layer further compares concrete path gain, execution risk, cost, and continuity with existing plans. This choose-strategy-then-choose-path design prevents the robot from oscillating among many similar candidates.

Quantitative Results: Generalization in Unknown Environments

To validate the method, the team conducted systematic experiments on the AiMDoom and Matterport3D (MP3D) benchmarks, focusing on cross-scene-distribution generalization.

AiMDoom: Across Simple, Normal, Hard, and Insane difficulty levels, FAME outperforms all baselines in final coverage rate and AUC. In the most challenging Insane setting, final coverage rises from NBP's 47.2% to 53.4% , a 6.2 percentage-point improvement.

MP3D (real indoor scans): FAME achieves a reconstruction completion rate of 82.67% , again leading the baselines.

Cross-dataset transfer (AiMDoom → MP3D): Without any fine-tuning, a model trained on AiMDoom reaches 68.45% reconstruction completion on MP3D, surpassing NBP's 61.81% by 6.64 percentage points.

Qualitative Demonstration: Keeping Multiple Options Open

The accompanying animation compares NBP (left) and FAME (right) starting from the same position in the same scene. NBP tends to persist along a single direction, ignoring other unexplored areas—even when adjacent rooms or corridors exist, the robot continues forward, leaving parts of the environment insufficiently reconstructed. In contrast, FAME simultaneously generates multiple possible exploration directions and selects a more suitable route, enabling the robot to cover areas that would otherwise be missed and producing a more complete final map.

Future Outlook

FAME's contribution goes beyond finding better exploration routes; it enables robots to model and imagine multiple future possibilities from limited observations and use those imaginings to guide current decisions . This aligns with the world-model paradigm of "understand the environment, imagine the future, guide action." The team plans to extend this direction from predicting future exploration directions toward imagining and reasoning about unknown scenes and action outcomes, allowing robots not only to perceive the world in front of them but also to make reasonable predictions about the unseen world, further advancing autonomous exploration and decision-making capabilities.

Paper Information

Title: All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping

Project Page: https://vollichor.github.io/FAME/

ArXiv: https://arxiv.org/abs/2609.36889

First Author: Li Yang, PhD student (2026) at Tianjin University School of Artificial Intelligence, advised by Prof. Ya-Hong Han.

Venue: NeurIPS 2026 Spotlight (top 3.7% of 7,900 accepted papers).