Source-linked AI summary
NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration
Ajay Sridhar, Dhruv Shah, Catherine Glossop, Sergey Levine
TL;DR
Robotic navigation in unfamiliar environments requires both finding goals through exploration and reaching goals once located, whereas these behaviors are often handled by separate models. NoMaD trains a unified diffusion policy with goal masking and action diffusion, and reports improved navigation performance with lower computational demands in unseen environments.
Problem
Navigation in unfamiliar environments requires both task-agnostic exploration and task-oriented goal-reaching, but these behaviors are typically handled by separate models.
Method
NoMaD uses goal masking with a Transformer-based policy and diffusion modeling of actions to support goal-conditioned navigation and undirected exploration.
Results
NoMaD consistently outperforms the tested baselines, producing smooth, reactive policies and requiring a 15× smaller model for exploratory goal discovery.
Takeaways & Limitations
A single goal-masked diffusion policy can perform both task-agnostic exploration and task-oriented navigation in previously unseen environments.
Takeaways & Limitations
The system uses goal images and standard frontier-based high-level exploration, limiting its current goal modalities and exploration-region selection.
Abstract
from arXiv · showhide
Robotic learning for navigation in unfamiliar environments needs to provide policies for both task-oriented navigation (i.e., reaching a goal that the robot has located), and task-agnostic exploration (i.e., searching for a goal in a novel setting). Typically, these roles are handled by separate models, for example by using subgoal proposals, planning, or separate navigation strategies. In this paper, we describe how we can train a single unified diffusion policy to handle both goal-directed navigation and goal-agnostic exploration, with the latter providing the ability to search novel environments, and the former providing the ability to reach a user-specified goal once it has been located. We show that this unified policy results in better overall performance when navigating to visually indicated goals in novel environments, as compared to approaches that use subgoal proposals from generative models, or prior methods based on latent variable models. We instantiate our method by using a large-scale Transformer-based policy trained on data from multiple ground robots, with a diffusion model decoder to flexibly handle both goal-conditioned and goal-agnostic navigation. Our experiments, conducted on a real-world mobile robot platform, show effective navigation in unseen environments in comparison with five alternative methods, and demonstrate significant improvements in performance and lower collision rates, despite utilizing smaller models than state-of-the-art approaches. For more videos, code, and pre-trained model checkpoints, see https://general-navigation-models.github.io/nomad/
I. INTRODUCTION
NoMaD addresses visual navigation in unfamiliar environments by unifying exploration and goal-directed behavior in one diffusion policy. The approach combines a Transformer visual backbone with diffusion-based action modeling for flexible task-specific and task-agnostic navigation.
- NoMaD targets navigation where robots must search unfamiliar environments for visually specified destinations.
- The method combines a Transformer backbone for visual observations with diffusion models that generate sequences of future actions.
- NoMaD jointly represents exploratory task-agnostic behavior and goal-directed task-specific behavior through one unified diffusion policy.
- Experiments cover undirected exploration and goal-conditioned navigation in challenging indoor and outdoor environments.
- NoMaD reports improved performance over state-of-the-art methods while using 15× fewer computational resources.
II. RELATED WORK
Prior exploration methods use local policies, global frontier strategies, privileged simulation information, reinforcement learning, or latent-variable planning. NoMaD is distinguished from closely related approaches by directly modeling actions with diffusion rather than generating subgoal images.
- Exploration methods use local control strategies or global frontier-based strategies to guide robots through environments.
- Simulation-trained exploration policies often transfer poorly to real-world environments.
- ViNT combines a goal-conditioned navigation policy with a separate 300M-parameter image diffusion model for subgoal proposals.
- Predicting exploratory action sequences is difficult because their distributions can be multimodal.
- Conditional diffusion models represent complex action distributions without inferring future states or observations.
III. PRELIMINARIES
The navigation policy maps current and past RGB observations, optionally paired with a goal image, to future actions. NoMaD combines goal masking, action diffusion, and topological-map memory for goal-reaching and long-horizon exploration.
- The policy takes current and past RGB observations as input and outputs a distribution over future actions.
- A goal image can specify the navigation task, requiring actions that progress toward and eventually reach the goal.
- NoMaD integrates its policy with episodic memory represented as a topological graph for long-horizon navigation.
- The architecture uses attention-based goal masking and a diffusion policy as its two key components.
- Action diffusion provides an expressive prior over collision-free actions for goal-reaching and undirected exploration.
A. Goal Masking
Goal masking controls whether the policy can use the goal-conditioned pathway. A binary mask blocks attention to the goal token for undirected behavior while preserving optional goal conditioning.
- The binary goal mask m determines whether the policy uses the optional goal-image pathway.
- Setting m = 1 masks the goal token so downstream context computation does not attend to it.
- Goal masking enables one shared policy to support both goal-reaching and undirected exploration.
B. Diffusion Policy
NoMaD uses diffusion to model complex, multimodal action distributions for both goal-conditioned navigation and goal-agnostic exploration. It iteratively denoises noisy action sequences while conditioning on visual context that may include a goal.
- Policy design: Diffusion approximates the conditional action distribution p(a_t|c_t), enabling complex multimodal navigation behaviors after goal masking.Without a goal, the policy can assign high probability to multiple safe actions, such as left and right turns at a junction, while suppressing collision-prone actions.
- Policy behavior: NoMaD predicts bimodal collision-free actions without a goal and narrower action distributions after conditioning on either of two goal images.Figure 3 visualizes task-agnostic and goal-directed pathways using yellow, green, and blue trajectories.
- Denoising process: The policy samples noisy future action sequences and performs K denoising iterations until producing a noise-free action sequence.Intermediate sequences progress from aK_t through a0_t with decreasing noise.
- Denoising process: The iterative update uses a noise prediction network εθ, with α, γ, and σ determined by the noise schedule.The denoising-step index is k, and the schedule controls the update coefficients and noise term.
- Training objective: The noise prediction network is conditioned on observation context that may or may not include goal information, while the denoising output models only actions.This conditional formulation supports real-time control and end-to-end training of the diffusion process and visual encoder.
C. Training Details
NoMaD is trained end-to-end with supervised learning on heterogeneous real-world navigation data from multiple environments and robotic platforms. Training balances goal-reaching and undirected exploration through goal masking.
- Data and supervision: NoMaD is trained on GNM and SACSoN datasets containing over 100 hours of real-world trajectories across diverse environments and robotic platforms.The data include pedestrian-rich environments and multiple ground-robot platforms.
- Data and supervision: Goal masking uses probability p_m = 0.5, providing an equal number of goal-reaching and undirected-exploration samples.This trains the unified policy on both task-specific and task-agnostic behavior.
- Optimization: The diffusion policy uses K = 10 denoising steps, a Square Cosine Noise Scheduler, and a 1D conditional U-Net with 15 convolutional layers.Optimization uses AdamW, learning rate 10^-4, batch size 256, and 30 training epochs.
V. EVALUATION
The evaluation tests NoMaD in six indoor and outdoor environments against prior methods and examines the value of joint behavior modeling and visual design choices. The experiments also vary the visual encoder and goal masking.
- Evaluation questions: NoMaD is evaluated in 6 distinct indoor and outdoor environments for visual exploration and goal-reaching in real-world settings.The first evaluation question compares NoMaD with prior work across these navigation tasks.
- Evaluation questions: The evaluation compares a joint task-agnostic and task-specific policy with individual behavior policies.This directly tests whether combining exploration and goal-directed behavior benefits navigation.
- Evaluation questions: The experiments examine how visual encoder choice and goal masking affect NoMaD performance.These are treated as separate design factors in the evaluation questions.
A. Benchmarking Performance
Across challenging indoor and outdoor environments, NoMaD consistently outperforms or matches strong baselines for exploration and navigation while using substantially fewer parameters. Its predictions also support smooth, collision-avoiding, multimodal behavior.
- Example rollouts show successful exploration trajectories in challenging indoor and outdoor environments on the LoCoBot platform.
- NoMaD paired with a topological graph consistently outperforms all baselines for exploration in unseen environments and navigation in known environments.
- 25%: NoMaD outperforms the state-of-the-art Subgoal Diffusion baseline while avoiding collisions and requiring 15× fewer parameters.
- NoMaD succeeds in all but the hardest environment and exceeds Subgoal Diffusion by over 25% in both exploratory efficiency and collision avoidance.
- NoMaD matches the performance of the best individual behavior policies for undirected exploration and goal-conditioned navigation despite comparable model capacities.
- NoMaD consistently captures multimodal action distributions and predicts accurately when conditioned on a goal image, unlike less robust or largely unimodal baselines.
B. Unified v/s Dedicated Policies
Goal masking enables one NoMaD policy to perform both goal-conditioned navigation and undirected exploration. Despite comparable capacity to dedicated policies, the unified model matches their performance while representing multimodal action choices.
- The jointly trained NoMaD model is compared with the best-performing dedicated goal-conditioned and undirected policies using mean success rate.
- The dedicated Diffusion Policy is the best exploration baseline, while ViNT is the best navigation baseline.
- NoMaD consistently represents multimodal undirected predictions while avoiding obstacles and correctly predicts actions toward two different goal images.
- The ViNT encoder with attention-based goal masking outperforms alternative visual encoder and masking strategies.
- NoMaD matches ViNT for goal-conditioned navigation and Diffusion Policy for undirected exploration despite comparable model capacities.
C. Visual Encoder and Goal Masking
NoMaD’s visual encoder strongly affects diffusion-policy training and goal-reaching performance. The model outperforms CNN- and ViT-based alternatives while reaching goals without collisions.
- Visual encoder choice is crucial for training NoMaD’s diffusion policies.The paper summarizes this finding in Table III.
- NoMaD outperforms both ViT- and CNN-based architectures while reaching the goal and avoiding collisions.
- Early-fusion CNN outperforms late-fusion CNN but struggles to condition effectively on goal information.
- The ViT encoder struggles to learn a good policy, likely because end-to-end diffusion training creates optimization challenges.
VI. DISCUSSION
NoMaD unifies task-agnostic exploration and task-oriented navigation with a masked-attention Transformer and diffusion policy. In unseen indoor and outdoor environments, it improves performance while using substantially fewer computational resources, but remains limited in goal modalities and exploration planning.
- NoMaD performs both task-agnostic exploration and task-oriented navigation with a unified goal-conditioned diffusion policy.
- A high-capacity Transformer encoder with masked attention flexibly conditions the diffusion model on task information such as goal images.
- NoMaD required 15× fewer computational resources while being evaluated on long-horizon exploration and navigation in unseen indoor and outdoor environments.
- The current system uses goal images and standard frontier-based exploration, leaving language, spatial coordinates, and semantic region selection for future work.