💻 computer science

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

VANDERER is a map-free exploration framework that enhances pre-trained diffusion policies for sensor-constrained mobile agents by utilizing a Visual Curiosity Module to predict action outcomes and guide navigation toward maximizing unexplored area coverage.

Original authors: Venkata Naren Devarakonda, Raktim Gautam Goswami, Prashanth Krishnamurthy, Farshad Khorrami

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Venkata Naren Devarakonda, Raktim Gautam Goswami, Prashanth Krishnamurthy, Farshad Khorrami

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are dropped into a massive, unfamiliar city with only a smartphone camera in your hand. You have no GPS, no map, and no 3D scanner. Your only job is to walk around and see as much of the city as possible without getting lost or crashing into things.

This is the challenge the VANDERER paper tackles. It's about teaching a robot (or an autonomous agent) to explore the world efficiently using just a single camera, without needing expensive sensors like LiDAR or GPS.

Here is how the paper solves this problem, broken down with simple analogies:

1. The Problem: The "Map-Maker" Trap

Traditional robots try to build a perfect 3D map of the world as they walk. They need expensive hardware (like laser scanners) to do this accurately. If they don't have these tools, they get confused.

The Paper's Approach: Instead of trying to build a perfect map, VANDERER acts like a curious tourist who just wants to see new things. It doesn't care about the exact geometry of the street; it just wants to know, "Have I seen this view before?"

2. The Engine: The "Diffusion Policy" (The Creative Artist)

The paper uses something called a Diffusion Policy. Think of this as a highly trained artist who has seen millions of videos of cars driving safely.

How it works: If you ask this artist to "draw a driving path," they don't just guess randomly. They start with a messy scribble (noise) and slowly refine it into a smooth, safe driving path based on what they've learned.
The Limitation: Without extra help, this artist might just draw a path that goes back and forth in the same spot because they are just trying to be "safe," not "exploratory." They need a nudge to go somewhere new.

3. The Guide: The "Visual Curiosity Module" (The Scout)

This is the secret sauce of VANDERER. It acts like a Scout standing next to the Artist.

The Process:
1. The Artist suggests a few different paths (e.g., "Turn left," "Go straight," "Turn right").
2. The Scout uses a "World Model" (a mental simulator) to imagine what the camera would see if the robot took those paths.
3. The Scout then checks a "memory bank" of everything the robot has already seen.
4. The Curiosity Test: The Scout asks, "Is this new view different from what we've seen?"
  - If the view is very similar to old views, the Scout says, "Boring! Don't go there."
  - If the view is totally new, the Scout says, "Exciting! Go there!"

4. The Magic Trick: Guiding the Artist

Instead of just picking the single "best" path (which might be a trap), VANDERER uses the Scout's feedback to tweak the Artist's initial sketch.

The Analogy: Imagine the Artist is painting a picture. The Scout whispers, "Hey, the left side of the canvas looks too much like yesterday's painting. Let's shift the colors on the left to make it look more like a forest."
The Artist then redraws the path, but this time, the path naturally leans toward the "new" areas because the starting "noise" was adjusted to favor them.

5. The Results: More Ground Covered, Fewer Crashes

The paper tested this in a realistic city simulation (CARLA) with five different towns.

The Competition: They compared VANDERER against other top methods (like NoMaD).
The Outcome: VANDERER explored 13.4% more area than the best competitor.
Safety: While the competitor often crashed into walls or poles (because it was trying to be too precise without good sensors), VANDERER crashed very rarely.
Why? The "Curiosity" guide kept the robot from getting stuck in loops (driving in circles) or crashing. It kept the robot moving toward the "unknown."

Summary

VANDERER is like a robot explorer that doesn't need a map. Instead, it uses a creative AI to plan moves and a curious scout to imagine the future. The scout constantly checks, "Have I seen this before?" and if the answer is "No," it nudges the robot to go there. This allows the robot to explore huge, complex outdoor areas efficiently using nothing but a standard camera.

Technical Summary: VANDERER

Problem Statement

Autonomous mobile agents require efficient exploration strategies to map unseen environments and plan tasks. Traditional approaches rely on constructing global occupancy maps, which necessitate precise localization and high-fidelity sensors (e.g., LiDAR, IMUs, depth cameras). These hardware requirements impose significant constraints, particularly in sensor-constrained settings where agents are limited to monocular cameras. Furthermore, in outdoor environments, geometric gaps may be misidentified as navigable frontiers, and semantic or legal constraints (e.g., traffic regulations) often restrict movement regardless of geometric accessibility.

Existing learning-based methods, particularly those using Reinforcement Learning (RL), often require extensive environment-specific training and large datasets. While diffusion policies have emerged as a promising alternative for generalization, they typically rely on policy variance for discovery and lack the global information necessary for long-horizon exploration. A fundamental challenge remains: how to train large-capacity models to behave exploratively without relying on explicit expert exploration data, which is scarce and difficult to define as a single optimal behavior.

Methodology

The paper proposes VANDERER, a map-free exploration framework designed for outdoor environments using only monocular RGB data. The system employs a "generate-then-optimize" approach, leveraging a pre-trained diffusion policy guided by a Visual Curiosity Module (VCM).

Core Components

Diffusion Policy:
- The policy is trained to imitate trajectory data, predicting feasible and safe action sequences (waypoints in a 2D top-down projection) directly from RGB observations.
- It utilizes a Variational Autoencoder (VAE) to encode current and historical frames into latent states.
- A context transformer processes these states to condition a UNet-style diffusion model, which iteratively denoises a latent action sequence.
Visual Curiosity Module (VCM):
- Future State Prediction: The VCM utilizes a Navigation World Model (NWM) to predict future latent states resulting from candidate action sequences generated by the diffusion policy.
- Curiosity-Based Cost Function: To quantify the "exploratory value" of predicted states without global maps, the system compares the predicted future state against a database of previously observed states ( $S$ ). Novelty is assessed using fast reciprocal nearest neighbor matching on patch-level features (via MASt3R). The cost ( $L_c$ ) is defined as the minimum distance between the predicted state and the most similar state in the database. Lower costs indicate higher novelty.
- Diffusion Guidance: Instead of greedily selecting the best action, the VCM guides the diffusion process by optimizing the initial noise distribution. Using a sampling-based strategy (Cross-Entropy Method), the system evaluates multiple noise samples, selects the top-performing candidates based on the curiosity cost, and updates the mean and variance of the noise distribution. This steers the sampling process toward actions that maximize environmental coverage while maintaining generation stability.

Inference Strategy

During inference, the agent generates multiple action candidates from the diffusion policy. The VCM predicts the outcomes of these actions, calculates curiosity costs, and refines the noise distribution used for sampling. This allows the pre-trained policy to adapt to exploration objectives without additional training or specialized expert data.

Key Contributions

Novel Framework: An end-to-end framework for map-free outdoor exploration using a visual-curiosity-guided diffusion policy, eliminating the need for complex mapping and expensive sensors.
Visual Curiosity Module: A module that predicts action consequences via a world model and guides the diffusion policy to generate outputs with higher exploratory potential by optimizing the noise distribution based on novelty costs.
Inference-Stage Optimization: A strategy that adapts pre-trained diffusion policies for efficient exploration without requiring additional training or specialized exploration datasets.
Empirical Validation: Extensive experiments demonstrating that VANDERER outperforms state-of-the-art baselines in exploration efficiency.

Results

The method was evaluated in five diverse "Town" environments within the CARLA simulator, comparing VANDERER against established baselines: NoMaD (an RGB-only diffusion method), NoMaD* (a modified version of NoMaD with a similar architecture to ensure fair comparison), and DP-RND (using Random Network Distillation as a selection metric).

Exploration Efficiency: VANDERER consistently outperformed all baselines. Across five environments, it explored an average of 13.4% more area than NoMaD.
Area Per Path Length (APL): VANDERER achieved the highest APL scores, indicating more efficient coverage relative to the distance traveled.
Robustness: While NoMaD showed slightly higher coverage in some instances, it suffered from a significantly higher policy failure rate (collisions) compared to VANDERER. VANDERER maintained a low failure rate (approx. 0.046%) comparable to the modified NoMaD*.
Ablation Studies:
- Removing the fast reciprocal matching module resulted in a performance drop, confirming its necessity for handling geometric variances in visual similarity.
- Replacing the diffusion policy with a standard Gaussian distribution (relying solely on CEM optimization) led to suboptimal actions, frequent collisions, and a 4x increase in execution time, highlighting the importance of the diffusion prior.
- Comparisons with a greedy selection baseline showed that greedy approaches often lead to repetitive looping and oscillation, whereas VANDERER's guidance mechanism maintains necessary stochasticity for robust exploration.

Significance and Claims

The paper claims that VANDERER addresses the critical limitations of sensor-constrained exploration by demonstrating that visual curiosity is a powerful metric for driving exploration in the absence of global maps or odometry. The work establishes a direct correlation between visual and geometric curiosity in outdoor environments, proving that agents can effectively leverage this relationship for efficient exploration using only monocular cameras.

Furthermore, the authors highlight the potential of pre-trained diffusion models to be guided toward complex exploratory objectives without the need for specialized expert data. By optimizing the inference process rather than retraining the model, VANDERER offers a scalable solution for autonomous agents operating in large-scale, unmapped outdoor environments with limited hardware resources.