Source-linked AI summary
ODG-NoMaD: Overhead-Camera Direction-Guided NoMaD
Blossom Treesa Bastian, Keerthi S. Shetty, Manish Kolachalam, Rani Malhotra, Ashish Dutta
TL;DR
NoMaD explores unseen environments without global awareness when no goal image or map is available. ODG-NoMaD adds inference-time directional guidance from an overhead-camera global path and onboard traversability, reducing residual distance by roughly 7×, outperforming NaviDiffusor, and avoiding collisions on every trial. The approach is evaluated within bounded workspaces fully observed by a calibrated overhead camera.
Problem
NoMaD’s exploration mode is undirected in unseen environments because its onboard, egocentric observations provide no global situational awareness.
Method
ODG-NoMaD uses an overhead-camera map and global path to form a directional prior, then applies traversability-based selection to guide NoMaD at inference time without retraining.
Results
Roughly 7× lower mean residual distance, greater reliability than NaviDiffusor’s point-goal cost guidance, and collision-free success on every trial were reported across both simulated environments.
Takeaways & Limitations
ODG-NoMaD restores goal-directed exploration while retaining a learned policy’s reactive, collision-aware behavior and handling obstacles absent from the overhead map.
Takeaways & Limitations
The approach assumes a bounded workspace observed in full by a single calibrated overhead camera.
Abstract
from arXiv · showhide
NoMaD [31] is a learned vision-navigation policy that unifies goal-conditioned navigation and exploration in a single goal-masked diffusion policy. In an unseen environment, however - where neither a goal image nor a topological map is available - it can only explore undirectedly, wandering without global awareness. We present ODG-NoMaD, which gives NoMaD's exploration mode a global sense of where to proceed, without retraining the policy. An overhead depth camera is used once on deployment to build an occupancy map and plan a global path, which is segmented to yield a desired heading; a per-frame traversability map from the robot's onboard depth then refines this into a collision-free direction. The gradient of a cosine direction cost is injected into the final denoising steps, rotating sampled trajectories toward this direction while preserving the multimodality of exploration. In simulated office environments with and without random obstacles, ODG-NoMaD reduces the residual distance to the target by up to an order of magnitude over unguided exploration, outperforms the point-goal cost guidance of NaviDiffusor [37], and is the only configuration that remains collision-free on every trial.
1 Introduction
ODG-NoMaD addresses NoMaD’s undirected exploration in unseen environments by adding globally informed inference-time guidance without retraining. It combines an overhead-camera global path with onboard traversability guidance to steer exploration toward targets while retaining collision-aware behavior.
- From the same start and goal, unguided NoMaD wanders and halts far short, whereas ODG-NoMaD follows the global path along a markedly shorter route.
- NoMaD’s exploration mode samples locally diverse, feasible trajectories but lacks global situational awareness because onboard observations are egocentric and limited by field of view and occlusions.
- ODG-NoMaD introduces globally aware signals during inference, leaving the trained navigation policy untouched.
- An overhead-camera occupancy map supports one-time global planning, whose path defines a directional prior for reweighting NoMaD’s sampled trajectories toward the target.
- A traversability-map collision-avoidance scheme steers around obstacles that are out-of-distribution for the learned policy while preserving alignment with the global path.
2 Related Work
Related work combines external overhead sensing for global awareness with inference-time steering of diffusion navigation policies. ODG-NoMaD connects these threads by deriving guidance from a global overhead view rather than from human input or egocentric costs.
- External overhead sensing: Overhead and ceiling-mounted cameras provide shared-world localization, mapping, and planning information for mobile robots.
- External overhead sensing: Overhead viewpoints supply global free-space, obstacle, and pose information unavailable to first-person cameras because of limited field of view and occlusions.
- External overhead sensing: Prior overhead-sensing systems typically feed top-down information into classical localization and planning stacks rather than learned reactive policies.
- Diffusion trajectory steering: NoMaD unified goal-conditioned navigation and undirected exploration in a single onboard-RGB goal-masked diffusion policy, motivating later trajectory-steering methods.
- Diffusion trajectory steering: NaviDiffusor guides denoising with differentiable scene- and task-level cost gradients, while other methods use human directional cues or egocentric costs.
- ODG-NoMaD’s position: ODG-NoMaD combines overhead global awareness with inference-time NoMaD steering without retraining, unlike prior guidance supplied by humans or robot-centric costs.
3 Overhead-Camera Direction-Guided NoMaD
ODG-NoMaD builds a one-time overhead-camera occupancy map and global path, extracts a desired heading, and combines it with per-frame onboard traversability to guide NoMaD’s exploration trajectories. The resulting direction cost is injected during inference-time denoising while the pretrained policy remains frozen and multimodal exploration is preserved.
- 3.1 Map Construction and Global Planning: A one-time overhead depth frame is converted into a top-down occupancy grid, on which a grid-based planner computes a collision-free path from the robot’s position to the target.The map uses height-based free, occupied, and unknown labels at 0.05 m resolution.
- 3.2 Path Segmentation and Direction Extraction: RDP simplifies the dense global polyline into near-straight segments, whose headings provide the desired travel direction for the robot’s current segment.The tolerance ε trades path fidelity against the number of retained segments.
- 3.3 Direction- and Collision-Guided Trajectory Modification: NoMaD encodes onboard RGB observations and uses a goal-masked action-diffusion policy to sample multimodal trajectory candidates over K denoising steps.The exploration-mode policy operates without a goal image.
- 3.3 Direction- and Collision-Guided Trajectory Modification: Inference-time guidance rotates sampled trajectories toward the desired direction without retraining the policy or collapsing its multimodal spread.The diffusion model remains frozen, and the guidance is injected only during denoising.
- 3.3 Direction- and Collision-Guided Trajectory Modification: The direction cost measures cosine alignment between each metric trajectory step and the desired body-frame direction, with its gradient applied during the final denoising steps.Normalized actions are de-normalized into metric displacements, cumulatively converted into waypoint positions, and evaluated against the desired unit direction.
- 3.3.2 Traversability Mapping and Best-Direction Selection: Per-frame traversability selects a collision-free direction by combining hard admissibility with a soft score that penalizes angular deviation and unknown-space exposure.The selected direction is the admissible candidate of smallest angular deviation after accounting for near-range ray coverage and unknown cells.
4 Experiments and Results
Experiments compare unguided NoMaD, NaviDiffusor goal guidance, and two ODG-NoMaD variants in office environments with consistent and unmapped obstacles. Direction guidance improves goal-reaching, while traversability-based selection provides collision-free behavior around obstacles absent from the overhead map.
- Environment Setup: Experiments use a Gazebo office with a ceiling-mounted overhead depth camera and a TurtleBot 4 equipped with an OAK-D RGB-depth camera.RGB observations feed NoMaD, while the overhead camera provides the global view used to construct the map and path.
- Environment Setup: The two environments differ by whether randomly placed obstacles are introduced after the overhead map and global path are generated.Env-NoObs matches the mapped scene; Env-RandObs tests obstacles unavailable to the planner.
- Configurations: Four configurations share the same pretrained NoMaD checkpoint and differ only in inference-time guidance: unguided NoMaD, NaviDiffusor, and two ODG-NoMaD variants.The ODG-NoMaD variants isolate direction guidance and add traversability-based collision avoidance in the full method.
- Results: ODG-NoMaD configurations reach the goal on every run in both environments, reducing goal closeness and increasing SPL relative to unguided exploration.The directional prior from the overhead global path steers the policy without retraining.
- Results: NaviDiffusor succeeds on fewer runs and achieves substantially lower SPL than ODG-NoMaD despite receiving the same global path information.Its point-directed cost guidance can stall short of the goal or fail on demanding routes.
- Results: In Env-RandObs, full ODG-NoMaD remains collision-free while the direction-only ablation collides with unmapped obstacles, although both variants are collision-free in Env-NoObs.The full method is the only configuration that succeeds on every trial without a collision.
5 Conclusion
ODG-NoMaD restores goal-directed exploration without retraining by combining an overhead-camera global path with traversability-based local collision avoidance. It substantially improves target closeness over unguided exploration, remains reliable around newly introduced obstacles, and assumes a bounded workspace observed by one calibrated overhead camera.
- Contribution: ODG-NoMaD guides NoMaD toward the target using a global path from an overhead-camera map and a traversability-based selector for obstacle avoidance.The method operates at inference time without retraining the policy.
- Evaluation: Table 1 evaluates three start–goal pairs in Env-NoObs using success rate, goal closeness, collision rate, and SPL, with three runs averaged per pair.The table marks the best value for each path–metric combination.
- Results: Roughly 7× lower mean residual distance than unguided exploration was achieved across both simulated environments.The comparison spans environments with and without randomly inserted obstacles.
- Results: ODG-NoMaD remains on the global path while steering around obstacles inserted after the overhead map and path were computed.This setting tests obstacles absent from the initial overhead map.
- Scope: The approach assumes a bounded workspace fully observed by a single calibrated overhead camera.Several overhead views or a sweeping drone are proposed as possible extensions for larger areas.
- Future scope: Semantic map labels could extend target specification from metric coordinates to objects, enabling object-goal navigation.This is presented as a future extension rather than an evaluated capability.