💻 computer science

ODG-NoMaD: Overhead-Camera Direction-Guided NoMaD

ODG-NoMaD enhances the NoMaD vision-navigation policy by integrating overhead-camera-derived global direction guidance into its diffusion-based exploration process, significantly reducing target distance and ensuring collision-free navigation in unseen environments without requiring policy retraining.

Original authors: Blossom Treesa Bastian, Keerthi S. Shetty, Manish Kolachalam, Rani Malhotra, Ashish Dutta

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Blossom Treesa Bastian, Keerthi S. Shetty, Manish Kolachalam, Rani Malhotra, Ashish Dutta

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 a robot sent into a room it has never seen before, tasked with finding a specific spot on the floor. The robot has no map, no instructions, and no one to guide it. It can only see what is directly in front of its eyes through a single camera. In such a situation, a robot trained to explore often behaves like a person wandering a new city without a compass: it moves forward, turns when it hits a wall, and occasionally circles back to places it has already visited. It might eventually stumble upon the goal by pure luck, but more often, it wanders aimlessly, trapped in local corners of the room, unable to grasp the bigger picture of where it needs to go. This is the core challenge of autonomous navigation in unknown spaces: how to give a robot the local reflexes to avoid immediate obstacles while also providing the global awareness needed to reach a distant destination.

Researchers have developed powerful learning-based systems that allow robots to learn how to move by watching thousands of examples, rather than being programmed with strict rules. One such system, known as NoMaD, is remarkably good at two things: navigating to a specific target when shown a picture of it, and exploring an unknown area when no target is given. However, in its exploration mode, the system lacks a sense of direction. It generates many possible paths, but without a global guide, it cannot tell which direction leads toward the goal. It is reactive, meaning it responds to what it sees right now, but it is not strategic. To solve this, a team of researchers from Infosys and the Indian Institute of Technology Kanpur has created a new method that combines the robot's learned ability to move with a simple, one-time view of the room from above.

The new approach, called ODG-NoMaD, works by giving the robot a brief glimpse of the entire workspace from a ceiling-mounted camera before it starts moving. This overhead camera takes a single depth image, which measures the distance to objects, and uses it to build a top-down map of the room. This map shows where the walls are, where the furniture sits, and where the open floor is. Using this map, a computer planner draws a safe, straight-line path from the robot's starting point to the target. This path is then broken down into simple directions, telling the robot which way to head at any given moment. The crucial innovation is that this global direction is not used to replace the robot's own decision-making. Instead, it is gently injected into the robot's existing learning model while it is running. The robot still uses its own camera to see the floor and avoid immediate bumps, but it now has a subtle bias pulling it toward the correct heading.

The researchers tested this system in a simulated office environment, a space filled with tables, chairs, and bins, similar to what one might find in a real workplace. They compared the new method against the standard, unguided version of the robot and another system that tries to guide the robot by pointing directly at the goal. The results showed a dramatic difference. Without any guidance, the standard robot wandered off course and stopped far from the target in most attempts. The new method, however, successfully guided the robot to the goal in every single trial, cutting the average distance it fell short by roughly seven times compared to the unguided version. It also performed better than the system that pointed directly at the goal, because the new method guided the robot along a general direction rather than forcing it toward a single point, which allowed it to remain flexible and adaptable.

Perhaps the most significant finding was how the system handled surprises. In some tests, the researchers placed new obstacles in the room after the overhead map had been created and the path planned. These obstacles were not on the map, so the robot could not have known they were there. The standard guided robot, which relied only on the pre-planned path, would have tried to drive straight into these new objects. The full ODG-NoMaD system, however, used a second, real-time check. As the robot moved, it used its own onboard camera to scan the floor directly in front of it. If it detected a new obstacle blocking the planned path, it would instantly calculate a safe detour, steering around the object while still keeping its general direction toward the goal. This allowed the robot to reach the target without hitting anything, even when the environment changed after the initial map was made.

The study demonstrates that it is possible to give a learning-based robot a sense of global direction without having to retrain it or teach it new rules. By simply adding a one-time overhead view and a way to check for immediate obstacles, the researchers turned a wandering explorer into a purposeful navigator. The system works entirely while the robot is running, using the pre-trained model exactly as it was, but steering its choices with a gentle, global nudge. This suggests a practical path forward for robots that need to operate in dynamic, unknown spaces, such as disaster zones or busy warehouses, where they must be able to find their way to a goal while reacting safely to the unexpected changes of the real world.

Technical Summary: ODG-NoMaD

Problem Statement
Learning-based navigation policies, such as NoMaD, excel at generalizing across environments and reacting to local observations by unifying goal-conditioned navigation and open-ended exploration within a single diffusion model. However, in unseen environments where no goal image or topological map is provided, NoMaD's exploration mode is fundamentally undirected. Lacking global situational awareness, the policy relies solely on egocentric, reactive decisions constrained by the camera's limited field of view. This leads to inefficient exploration characterized by revisiting covered regions, wandering without clear progress, and failing to reach targets. Conversely, classical planning methods provide global awareness but lack the reactivity and generalization of learned policies. The challenge lies in combining the global guidance of classical planning with the reactivity of learned exploration without retraining the policy or sacrificing its multimodal capabilities.

Methodology: ODG-NoMaD
The authors propose ODG-NoMaD (Overhead-Camera Direction-Guided NoMaD), an inference-time guidance framework that steers NoMaD's goal-masked diffusion policy toward a target using a global path derived from an overhead camera, without modifying the trained network weights. The pipeline operates in three stages:

Global Path Planning via Overhead Camera:
Upon deployment, a single overhead depth camera captures the workspace to construct a top-down occupancy grid map. A global planner computes a collision-free path from the robot's start position to the target. This dense path is simplified using the Ramer–Douglas–Peucker (RDP) algorithm into a sequence of straight segments. The segment containing the robot's current position provides a desired global heading ( $\theta^\star$ ) and unit direction vector ( $\hat{d}$ ).
Direction Guidance via Gradient Injection:
To bias the undirected exploration samples toward the global heading, the method injects the gradient of a direction cost function ( $F_{dir}$ ) into the denoising process of the diffusion policy during the final, low-noise steps.
- The cost function measures the cosine alignment between the predicted trajectory steps and the desired unit direction $\hat{d}$ .
- Crucially, the use of cosine alignment ensures the guidance rotates the trajectory steps toward the desired heading without artificially inflating the step lengths, thereby preserving the multimodal nature of the exploration.
- This process is training-free; the diffusion model remains frozen, and the global context is injected solely through the gradient update.
Traversability-Based Collision Avoidance:
To handle obstacles not present in the initial overhead map (out-of-distribution obstacles), the system generates a per-frame Bird's-Eye-View (BEV) traversability map using the robot's onboard depth camera.
- A "best-direction selector" evaluates a fan of candidate headings around the desired global direction.
- A hard rule rejects any candidate that intersects an obstacle within a near-range safety margin (including side rails to ensure lateral clearance).
- A soft rule scores admissible candidates based on angular deviation from the desired heading and the fraction of unknown space in the path.
- The selected direction ( $\hat{d}_{trav}$ ) is the admissible candidate with the smallest angular deviation, ensuring the robot avoids new obstacles while remaining aligned with the global path. This direction is then used as the target for the direction guidance cost.

Key Contributions

ODG-NoMaD Framework: A method that guides NoMaD's goal-masked diffusion policy at inference time using a directional prior derived from an overhead-camera global path. This steers exploration toward a target without human-in-the-loop intervention or policy retraining.
Traversability-Based Collision Avoidance: A scheme that integrates per-frame onboard depth data to select a collision-free heading. This allows the robot to deviate around new, unmapped obstacles while maintaining alignment with the global path.
Comparative Analysis: A head-to-head comparison demonstrating that directional guidance is more reliable for reaching the final point of a global path than the point-goal cost guidance used in NaviDiffusor. The authors attribute this to the fact that a heading prior preserves the policy's exploratory multimodality, whereas a precise point target collapses the action distribution.
Robustness: The system demonstrates robustness to dynamic environments, successfully avoiding obstacles introduced after the overhead map is built while still reaching the goal.

Experimental Results
Experiments were conducted in a simulated Gazebo office environment (9.2m x 5.2m) using a TurtleBot 4. Two scenarios were tested: Env-NoObs (static furniture matching the map) and Env-RandObs (with random obstacles added post-mapping).

Performance vs. Unguided Exploration: ODG-NoMaD reduced the mean residual distance to the target by approximately 7× compared to unguided NoMaD exploration. While unguided NoMaD rarely reached the target (Success Rate ~33% in static, 0% in dynamic), ODG-NoMaD achieved a 100% success rate in both environments.
Performance vs. NaviDiffusor: ODG-NoMaD outperformed NaviDiffusor (which uses point-goal cost guidance) in terms of Success Rate and Success Weighted by Path Length (SPL). NaviDiffusor frequently stalled short of the goal or failed on complex routes, whereas ODG-NoMaD reached the target more reliably.
Collision Avoidance: In the Env-RandObs scenario, the version of ODG-NoMaD without traversability-based selection collided with new obstacles (100% collision rate), whereas the full method remained collision-free (0% collision rate) by deviating around them.
Metrics: The full ODG-NoMaD configuration achieved the highest Success Rates and SPLs across all paths and environments, with goal closeness significantly lower than all baselines.

Significance and Claims
The paper claims that ODG-NoMaD successfully restores goal-directed behavior to a learned, goal-masked diffusion policy in unseen environments. By leveraging a global overhead view for planning and an onboard view for local obstacle avoidance, the method bridges the gap between classical global planning and reactive learned exploration. The authors emphasize that this is achieved entirely at inference time, preserving the pre-trained policy's generalization capabilities while adding necessary global awareness. The approach is presented as a robust solution for mobile robots operating in partially known or dynamic environments where explicit map construction is difficult, but an overhead view is available for initial global context.

Technical Summary: ODG-NoMaD

More like this