NeurIPS 2026 Spotlight｜Why Can't Robots Just Follow One Path? Tianjin University Team Proposes FAME

The first author of this paper is Li Yang, a doctoral student at the School of Artificial Intelligence, Tianjin University, class of 2026, supervised by Professor Han Yahanong, with main research direction in embodied intelligence. Related research results have been published in international top conferences and journals such as CVPR, NeurIPS, and TPAMI. This study was completed through collaboration between the teams of Tianjin University and Hefei University of Technology. Lab homepage:

https://tjumm.github.io/

Imagine a robot exploring a building that it has never entered before.

It has walked down a corridor, and now several different directions of passageways appear ahead. However, due to limited vision, the robot does not know which path leads to a larger space, which might be a dead end, or what structure the unexplored area might have.

At this point, where should it go?

Currently, many learning-based active 3D mapping methods often only predict a single long-term exploration goal and plan a path based on that.

The issue is:

Why must the robot bet on a single direction when it does not yet know much about the unknown environment?

To address this issue, the research team from Tianjin University and Hefei University of Technology proposed

FAME（Flow-driven Anchor-based Multimodal Exploration）

. Unlike traditional methods that only predict a single target, FAME utilizes Conditional Flow Matching to generate multiple possible exploration targets simultaneously, then combines path planning, exploration pattern clustering, and hierarchical selection to help the robot find more suitable exploration routes in unknown environments.

Currently, this research result has been

accepted by NeurIPS 2026 and further selected as a Spotlight (highlight paper)

. This conference has accepted 7,900 papers, among which

292 papers

were selected as Spotlight, accounting for about 3.7%.

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

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

Paper link: https://arxiv.org/abs/2609.36889

From 'Only One Path' to Keeping Multiple Exploration Possibilities

Active 3D Mapping (Active 3D Mapping)

It not only requires robots to recover the three-dimensional structure of the environment based on observations, but also requires them to actively decide where to go next to collect information, striving to reconstruct as complete a scene as possible with fewer movements and observations.

However, the robot can only observe part of the environment.

Even if the local structures of two scenes appear similar, the unexplored areas may be completely different.

If the model predicts a long-term goal based solely on past learned experiences, it is prone to making wrong directional choices and missing other valuable areas when encountering unfamiliar environments.

The research team pointed out that, under partially observable conditions, long-term exploration goals often have

multi-modal features

. In other words, for the same current state, there may be multiple reasonable future exploration directions.

Theoretically, when multiple reasonable exploration goals are distributed in different directions, a single-target prediction model trained with squared error tends to give the

average position

of these goals. However, this position may not be a suitable exploration target, and may even fall within an obstacle area.

Therefore, Rather than letting the model predict only one target, it is better to let it consider multiple possible exploration directions simultaneously.

Figure 1: Motivation comparison between single-target prediction and multi-modal exploration

Generating multiple exploration hypotheses with Flow Matching

To enable the robot to generate multiple possible exploration targets simultaneously, the research team introduced Conditional Flow Matching（CFM）.

Specifically, the model generates multiple different positions and orientations of exploration targets, starting from random noise, based on the map the robot has already built. These are called

Anchor (anchor points)

. Each Anchor represents a potential long-term exploration direction.

For example, some anchors may guide the robot to enter a room on the left, some may correspond to continuing into the corridor, and others may point to an area that has not yet been covered.

Unlike directly generating a complete exploration route, FAME

only needs to predict the position and direction of the target

. In short, the model is responsible for suggesting several possible places to go, and the path planner then calculates how to get to each target based on the obstacle map.

This

allows for retaining multiple exploration directions

without requiring the model to directly generate the entire route.

Figure 2: Schematic diagram of multi-modal target generation with Conditional Flow Matching

It's not just about generating multiple routes; it's also about learning how to choose

As mentioned earlier, CFM can help the robot generate multiple exploration targets. However, having these targets is not enough; the robot also needs to determine which routes are feasible and which one to ultimately choose.

For this purpose, FAME uses

a path planner based on A*

to generate candidate paths for each Anchor based on the current obstacle map.

Subsequently, the system clusters similar paths based on information such as the endpoint of the path, geometric shape, initial execution behavior, and obstacle avoidance topology.

This way, even if multiple Anchors ultimately correspond to similar routes, they can

be organized into the same exploration mode, reducing the interference of redundant candidates on subsequent decisions.

Finally, there is hierarchical selection.

FAME does not sort all paths at once, but rather first makes a selection at the exploration mode level, and then filters specific routes within the selected mode.

The first layer focuses on potential exploration gains, risk, uncertainty, cost, and diversity; the second layer further compares specific paths in terms of gains, execution risk, cost, and continuity with existing planning.

This

The design of first selecting an exploration strategy and then selecting an execution route

enables the robot to avoid repeatedly switching between a large number of similar candidates.

Figure 3: Overall architecture diagram of FAME

Quantitative results: How well does the method generalize in unknown environments?

To test the effectiveness of the method, the research team conducted systematic experiments on two benchmarks, AiMDoom and Matterport3D (MP3D), with a focus on cross-scenario generalization capability.

On the AiMDoom dataset, FAME outperformed other comparison methods in terms of final coverage and AUC across four difficulty levels: Simple, Normal, Hard, and Insane. Among them, in the most challenging Insane setting, the final coverage increased from NBP's 47.2% to

53.4%, an improvement of 6.2 percentage points.

Table 1: Evaluation results on the AiMDoom dataset, where FAME's final coverage and AUC are superior to comparison methods across four difficulty levels

On the Matterport3D benchmark built from real indoor scan data, FAME achieved a reconstruction completion rate of

82.67%, which is also better than other comparison methods. More notably is the cross-dataset generalization capability. The research team trained the model on AiMDoom and directly transferred it to MP3D testing without any fine-tuning. In this more challenging setting, FAME still achieved a reconstruction completion rate of 68.45%, which is 6.64 percentage points higher than NBP's 61.81%.

Table 2, Table 3: Comparison of MP3D reconstruction completion rate and reconstruction error (left); cross-dataset transfer results from AiMDoom→MP3D (right), FAME achieves the best performance

Qualitative Demonstration: How Robots Can ``Leave More Paths'' in Unknown Environments

The following animated graphic shows NBP (left) and FAME (right) exploration processes in the same scenario and starting point.

It can be seen that

NBP tends to explore continuously in one direction, while ignoring other unexplored areas.

Even if there are rooms or corridors nearby, the robot may continue along the current path, leading to some areas not being adequately reconstructed.

In contrast, FAME does not focus on a single exploration goal, but instead generates multiple possible exploration directions and selects the more suitable path from them.

In this way, the robot has the opportunity to explore more areas that are otherwise easily overlooked, resulting in a more complete final reconstructed map.

Figure 4: NBP (left) versus FAME (right) exploration coverage comparison in the same scenario

Future Outlook

The research on FAME is not only about enabling robots to find more suitable exploration paths, but more importantly, enabling robots to model and imagine multiple possible futures based on limited observations, and using this to guide current decision-making.

This approach also aligns with the world model's emphasis on ``understanding the environment, imagining the future, and guiding actions``.

In the future, the research team will continue to explore in this direction, gradually expanding from predictions about future exploration directions to imagining and reasoning about unknown scenarios and action outcomes, enabling robots not only to perceive the world around them but also to make reasonable predictions about the world they have not yet seen, thereby further enhancing their autonomous exploration and decision-making capabilities.

© THE END

For republication, please contact this official account for authorization

For submissions or media inquiries: [email protected]