VANDERER: Map-Free Exploration using Future-Aware and Visual-Curiosity-Guided Diffusion Policy

arXiv:2606.14879 · cs.RO, cs.CV, cs.LG · Submitted 2026-06-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "VANDERER: Map-Free Exploration using Future-Aware and Visual-Curiosity-Guided Diffusion Policy".

Dev: Here's the gist of the paper. Mobile agents require efficient exploration strategies to map unseen environments and autonomously plan tasks <ref:2606.14879#pg2>.

Rosa: To address this, we propose VANDERER, an exploration framework that leverages a Visual Curiosity Module (VCM) to guide pre-trained diffusion policies using only monocular image data <ref:2606.14879#pg2>.

Dev: Pruning redundant reasoning tokens The paper introduces VANDERER, an exploration framework designed for efficient outdoor exploration that foregoes the overhead of complex mapping <ref:2606.14879#pg2>.

Rosa: This approach leverages a diffusion policy to generate feasible candidate low-level actions directly from RGB observations <ref:2606.14879#pg2>.

Paper summary: Rosa: So we're looking at this paper called "VANDERER: Map-Free Exploration using Future-Aware and Visual-Curiosity-Guided Diffusion Policy". The main idea is that mobile agents need better ways to explore unseen environments without needing to build those big occupancy maps, which usually takes a lot of computational power.

Dev: Right, so the whole thesis is this framework uses something called a Visual Curiosity Module or VCM to guide pre-trained diffusion policies using only monocular image data. It’s trying to figure out how an agent can explore efficiently without mapping everything out first <ref:2606.14879#pg2>.

Taro: So it’s about letting the system predict what happens next, and then using a curiosity cost function to decide which direction looks most interesting, rather than just following some pre-set path <ref:2606.14879#pg5>.

Rosa: Exactly. They're foregoing the overhead of complex mapping by letting a diffusion policy generate feasible low-level actions directly from the RGB observations <ref:2606.14879#pg3>.

Dev: And they build that training signal using a "generate-then-optimize" strategy, where the diffusion policy is trained to imitate actual trajectory data to make safe sequences of actions <ref:2606.14879#pg4>.

Taro: When it comes to inference, the guidance mechanism refines the noise distribution within the reverse diffusion process based on those curiosity costs <ref:2606.14879#pg5>. That means it steers the sampling process toward areas that haven't been seen much <ref:2606.14879#pg4>.

Rosa: The core of this guidance comes from that visual curiosity, which has three parts: predicting future states using a world model, having a curiosity-based cost function to measure novelty, and then applying that loss to guide the diffusion process <ref:2606.14879#pg5>.

Dev: That cost function quantifies exploratory value by comparing those predicted future states against a database of prior observations using fast reciprocal matching on patch level features <ref:2606.14879#pg4>. It’s basically measuring how surprising the next spot is compared to what the agent has seen before.

Taro: So this mechanism steers the sampling process toward unexplored areas, which sounds like it bypasses those greedy selection limitations we sometimes see in other methods <ref:2606.14879#pg6>.

Rosa: The results show that VANDERER consistently outperforms established baselines, exploring an average of thirteen point four percent more area than NoMaD one <ref:2606.14879#pg2>. That’s a significant gain when you have limited sensors, like just a single camera <ref:2606.14879#pg2>.

Dev: It sounds like the paper shows that visual curiosity is a powerful metric to drive exploration in sensor constrained environments <ref:2606.14879#pg2>. But what about the practical side? How does this actually run outside of a perfect simulation?

Taro: The ablation studies confirm that both the fast reciprocal matching and the diffusion policy are critical to its success, so you can't just pick one part out <ref:2606.14879#pg9>.

Rosa: That’s interesting because it means we have to keep both components working well for this method to hold up in the real world <ref:2606.14879#pg9>. It shows the interplay between the model and the guidance is important.

Dev: So, if you think about it, this approach replaces complex mapping with a guided diffusion process driven by curiosity, using only what a camera sees <ref:2606.14879#pg1>. It’s a way to get better coverage without needing the full map first.

Taro: What I see is that it leans into using pre-trained diffusion models to explore complex objectives without needing specialized expert data for every single scenario <ref:2606.14879#pg2>. That’s a big thing for general applicability.

Rosa: It suggests that if we can guide these powerful pre-trained models with just curiosity, we might be able to explore things in environments where we don't have the ground truth map data <ref:2606.14879#pg9>. But the paper does mention that this relies on a navigation world model predicting future states <ref:2606.14879#pg3>.

Dev: That world model is key, and it’s what allows the curiosity cost function to actually work by predicting outcomes of proposed actions <ref:2606.14879#pg3>. But we have to be careful about that prediction accuracy in real-time loops <ref:2606.14879#pg3>.

Taro: I’m concerned about the failure modes, though. If the world model makes a bad prediction about the future state, how gracefully does VANDERER recover when it's supposed to be exploring an unknown area? <ref:2606.14879#pg3>.

Rosa: That’s a fair question for real-world deployment. The paper focuses on the guidance mechanism steering toward unexplored areas, but it doesn't detail the specific safety case for when the world model is wrong <ref:2606.14879#pg3>.

Dev: It seems like VANDERER is really about coupling a diffusion policy with this visual curiosity guidance to achieve efficient exploration using just camera data <ref:2606.14879#pg9>. It’s a clever way to handle the sensor constraints of outdoor navigation.

Taro: And it proves that visual curiosity isn't just some fancy metric; it’s a concrete way to drive exploration when you only have limited sensing resources <ref:2606.14879#pg2>.

Rosa: So, in conclusion, VANDERER proposes a framework that uses visual curiosity to guide diffusion policies for map-free exploration using monocular images <ref:2606.14879#pg1>. It shows that pre-trained models can be guided toward complex exploratory objectives without needing extensive expert data <ref:2606.14879#pg2>.

Dev: And it certainly suggests that for mobile agents in constrained settings, focusing on visual novelty instead of a full map might be a much more efficient path to coverage <ref:2606.14879#pg9>.

Conclusion: Rosa: So we're wrapping up this look at VANDERER, which is about map-free exploration using diffusion policies guided by visual curiosity.

Dev: Right, so basically, this framework lets mobile agents explore an area without building a detailed map first by using those visual cues to guide the AI's decisions.

Rosa: The authors are focusing on how this works with just monocular images, which means they aren't relying on that extra processing power for mapping stuff.

Dev: And the core idea is using a diffusion policy trained to follow past paths and then tweaking that process during exploration based on how curious it finds new things.

Rosa: So what does this actually mean for us out here in the real world? It suggests we can get better coverage just by looking at pictures and being curious about what's ahead.

Dev: But we gotta remember the technical side, like how fast that guidance mechanism needs to run, and how robust it is when the camera sees something totally unexpected.

Rosa: Exactly. The paper shows visual curiosity is a powerful way to drive exploration in environments where you have limited sensors and computational resources.

Dev: It’s certainly a strong result because they showed it exploring about thirteen point four percent more area than this NoMaD method we talked about earlier in the show.

Rosa: That's a solid number, but the real test is seeing if that curiosity actually translates into useful exploration in unpredictable, messy outdoor settings.

Dev: We’ll be looking at how this framework handles those tricky moments when the world doesn't behave exactly as expected during those long-term planning steps.

Venkata Naren Devarakonda, Raktim Gautam Goswami, Prashanth Krishnamurthy, Farshad Khorrami

NYU Tandon School of Engineering

cs.RO, cs.CV, cs.LG

Submitted: 2026-06-12

Updated: 2026-11-13

Comments: 11 pages, 8 figures. Project video: https://youtu.be/u1z1SVjK4_A

Project page: https://narendv.github.io/vanderer

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Mobile agents require efficient exploration strategies to map unseen environments and autonomously plan tasks <ref:2606.14879#pg2>.

Key concepts

Visual Curiosity Module (VCM)
The VCM uses a navigation world model to predict what the agent will see in the future. It then calculates a curiosity-based cost function that measures how novel or surprising these predicted future states are compared to past observations, quantifying the potential exploratory value of a location.
Diffusion Policy
This is a pre-trained policy used by VANDERER to generate feasible low-level actions directly from RGB images. Instead of relying on complex mapping, this policy is trained to imitate successful trajectory data and can be guided during inference to favor exploratory behaviors.
Curiosity-Based Cost Function
This function quantifies the 'curiosity' of a future state by comparing it against a database of previously seen observations. If the predicted future state is significantly different from what has been seen before, it receives a higher curiosity score, which then influences the guidance mechanism.
Generate-then-Optimize Strategy
The training involves first generating action sequences using the diffusion policy to imitate existing data. During inference, this strategy is refined by optimizing the initial noise latent distribution within the reverse diffusion process. This optimization uses curiosity costs to steer sampling toward areas deemed more exploratory.

Terminology

Summary

Mobile agents require efficient exploration strategies to map unseen environments and autonomously plan tasks <ref:2606. p. 0

Where the paper says it

VANDERER: Map-Free Exploration using Future-Aware and Visual-Curiosity-Guided Diffusion Policy Venkata Naren Devarakonda1 , Raktim Gautam Goswami1 , Prashanth Krishnamurthy1 , Farshad Khorrami1 Abstract— Mobile agents require efficient exploration strategies to map unseen environments and autonomously plan tasks. Traditional methods rely on generating occupancy maps and optimizing the sequence in which unexplored regions are visited.

14879#pg2>. To address this, we propose VANDERER, an exploration framework that leverages a Visual Curiosity Module (VCM) to guide pre-trained diffusion policies using only monocular image data <ref:2606. p. 0
Where the paper says it

However, in sensor-constrained settings, such as those limited to monocular cameras, generating accurate occupancy maps is challenging. To address this, we propose VANDERER, an exploration framework that leverages a Visual Curiosity Module (VCM) to guide pre-trained diffusion policies using only monocular image data.

14879#pg2>.

Pruning redundant reasoning tokens

The paper introduces VANDERER, an exploration framework designed for efficient outdoor exploration that foregoes the overhead of complex mapping <ref:2606.14879#pg2>. This approach leverages a diffusion policy to generate feasible candidate low-level actions directly from RGB observations <ref:2606. p. 0

Where the paper says it

1), a framework designed for efficient outdoor exploration that foregoes the overhead of complex mapping. Our approach leverages a diffusion policy to generate feasible candidate low-level actions directly from RGB observations.

14879#pg2>. Central to this method is the Visual Curiosity Module (VCM), which uses a navigation world model to predict future states and evaluates their novelty via a curiosity-based cost function <ref:2606. p. 0
Where the paper says it

Central to our method is the Visual Curiosity Module (VCM), which utilizes a navigation world model [8] to predict future states resulting from candidate actions. The VCM quantifies the “exploratory value” of these predicted states by comparing them against a database of prior observations via a curiosity-based cost function.

14879#pg4>.

How the training signal is built

The VANDERER framework employs a “generate-thenoptimize” strategy where the diffusion policy is trained to imitate trajectory data to predict feasible and safe action sequences <ref:2606.14879#pg4>. During inference, the model is guided toward exploration-friendly behaviors by optimizing the initial noise latent distribution within the reverse diffusion process <ref:2606. p. 2

Where the paper says it

Our diffusion policy is trained to imitate trajectory data to predict feasible and safe action sequences. During inference, we guide the model toward exploration-friendly behaviors by optimizing the initial noise latent distribution within the reverse diffusion process.

14879#pg4>. This guidance mechanism refines the mean and variance of the noise distribution based on curiosity costs <ref:2606. p. 1
Where the paper says it

(a) A simplified example of action predictions generated from two noise latents sampled from a time-invariant Gaussian distribution through a pretrained diffusion policy. The curiosity guidance mechanism predicts the resulting future states and estimates their novelty against a database of prior observations.

14879#pg5>.

Where the gains come from

The core of VANDERER is its visual curiosity guidance, which comprises three components: future state prediction via a world model, a curiosity-based cost function, and the application of this loss to provide diffusion guidance <ref:2606.14879#pg5>. The curiosity cost quantifies exploratory value by comparing predicted future states against a database of prior observations using fast reciprocal matching on patch-level features <ref:2606. p. 0

Where the paper says it

Central to our method is the Visual Curiosity Module (VCM), which utilizes a navigation world model [8] to predict future states resulting from candidate actions. The VCM quantifies the “exploratory value” of these predicted states by comparing them against a database of prior observations via a curiosity-based cost function.

14879#pg4>. This mechanism steers the sampling process toward unexplored areas <ref:2606. p. 1
Where the paper says it

The curiosity guidance mechanism predicts the resulting future states and estimates their novelty against a database of prior observations. These novelty scores are then used to update the diffusion policy’s noise distribution, guiding the agent toward unexplored areas.

14879#pg4>.

Performance and validation

Extensive experimentation shows VANDERER consistently outperforms established baselines, exploring an average of 13.4% more area than NoMaD <ref:2606. p. 0

Where the paper says it

Evaluated across diverse simulated environments, VANDERER consistently outperforms established baselines, exploring an average of 13. 4% more area than NoMaD [1].

14879#pg2>. The method's guidance mechanism avoids the limitations of greedy selection by updating the noise distribution rather than greedily selecting the best candidate <ref:2606.14879#pg6>. Ablation studies confirm that both fast reciprocal matching and the diffusion policy are critical to its success <ref:2606. p. 6
Where the paper says it

To evaluate the contribution of the individual components of VANDERER, we conduct ablation studies by systematically removing key architectural elements. Having previously evaluated the impact of the VCM, we now assess the importance of fast reciprocal matching and the diffusion policy.

14879#pg9>. The results demonstrate that visual curiosity is a powerful metric to drive exploration in sensor-constrained environments <ref:2606. p. 7
Where the paper says it

Rather than employing hard selection, we propose a guidance strategy to better balance exploration and exploitation, resulting in superior APL. Through extensive comparisons, we substantiate that visual curiosity is a powerful metric to drive exploration in sensor-constrained environments.

14879#pg2>.

Conclusion

In settings with limited sensing and computational resources, efficient autonomous exploration remains a significant challenge <ref:2606.14879#pg9>. VANDERER presents a framework that couples a diffusion policy with a visual-curiosity-based guidance mechanism to enable efficient exploration using only camera data <ref:2606. p. 7

Where the paper says it

CONCLUSIONS In settings with limited sensing and computational resources, efficient autonomous exploration remains a significant challenge. To address this, we presented VANDERER, a framework that couples a diffusion policy with a visualcuriosity-based guidance mechanism to enable efficient exploration using only camera data.

14879#pg9>. The work substantiates that visual curiosity is a powerful metric to drive exploration in sensor-constrained environments <ref:2606. p. 7
Where the paper says it

Through extensive comparisons, we substantiate that visual curiosity is a powerful metric to drive exploration in sensor-constrained environments. Furthermore, our work highlights the potential for pre-trained diffusion models to be guided toward complex exploratory objectives without the need for specialized expert data.

14879#pg2>. Furthermore, the research highlights the potential for pre-trained diffusion models to be guided toward complex exploratory objectives without the need for specialized expert data <ref:2606. p. 7
Where the paper says it

Through extensive comparisons, we substantiate that visual curiosity is a powerful metric to drive exploration in sensor-constrained environments. Furthermore, our work highlights the potential for pre-trained diffusion models to be guided toward complex exploratory objectives without the need for specialized expert data.

14879#pg2>.

REFERENCES

arXiv:2606.14879v1 [cs. p. 0

Where the paper says it

arXiv:2606. 14879v1 [cs.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. A novel end-to-end framework for map-free exploration using a visual curiosity-guided diffusion policy allows agents to generate feasible candidate low-level actions directly from RGB observations without requiring occupancy maps or high-fidelity sensors like LiDAR or IMUs.

  2. The Visual Curiosity Module (VCM) utilizes a navigation world model to predict future states and quantifies the exploratory value of these predicted states by comparing them against a database of prior observations via a curiosity-based cost function, guiding the diffusion process toward generating actions that maximize exploration.

  3. The inference-stage optimization strategy adapts pre-trained diffusion policies for efficient exploration without the need for additional training, improving performance over state-of-the-art methods by consistently exploring an average of 13.4% more area in simulated environments compared to NoMaD.

  4. The system can effectively leverage the correlation between visual and geometric curiosity in outdoor environments, demonstrating that it can efficiently explore using only a monocular RGB stream, addressing hardware constraints common in sensor-constrained settings.

  5. The diffusion policy is conditioned on 3D feature tokens derived from VAE encoders and processed by a context transformer to generate action sequences, enabling the model to learn both navigation and exploration within a single architecture.

Abstract

Mobile agents require efficient exploration strategies to map unseen environments and autonomously plan tasks. Traditional methods rely on generating occupancy maps and optimizing the sequence in which unexplored regions are visited. However, beyond requiring additional sensor suites for reliable pose and depth estimation, occupancy map-based exploration policies often neglect environmental traversability constraints. To address these limitations, we propose VANDERER, an exploration framework that leverages a Visual Curiosity Module (VCM) to guide pre-trained diffusion policies using only monocular image data. This curiosity module predicts the outcomes of proposed actions via a navigation world model and evaluates them through a curiosity cost. The cost then guides the diffusion process toward generating actions that maximize exploration. Evaluated across diverse simulated environments, VANDERER outperforms established baselines, achieving 2.3 times the exploration efficiency of Frontier Search. Our results reveal a correlation between visual and geometric curiosity in outdoor environments, demonstrating that VANDERER can effectively leverage this relationship for efficient exploration using simple robotic setups.

Sources

Related papers