OPEN MODELS/NAVIGATION/NOMAD
INTERMEDIATENAVIGATION

NoMaD

Diffusion-based navigation that explores before it navigates

WHAT DOES THIS ACTUALLY DO?

Most navigation models need a known map. NoMaD doesn't — it first explores to build understanding, then navigates to goals. It's like the difference between wandering around a new city (exploring) vs. walking somewhere once you know the layout.

PARAMETERS
~30M
RELEASED
2024-01
LICENSE
MIT
ARCHITECTURE
Vision...
CAN I ACTUALLY RUN THIS?
7
Intermediate
OUT OF 10

Shares the ViNT codebase, so setup is similar. Good documentation and Jetson compatibility make this very accessible for mobile robotics developers.

DOWNLOAD WEIGHTS
View on GitHub
2GB disk space required
OVERVIEW

Goal Masking Diffusion Policies unifying task-agnostic exploration and goal-directed navigation in a single model. Superior performance in unseen environments.

TRAINED ON

GNM dataset + diverse exploration data

ARCHITECTURE
Vision Encoder + Diffusion Policy + Goal Masking
CAPABILITIES
Unified exploration and navigation
Diffusion-based action generation
Goal masking for flexible conditioning
Novel environment handling
REAL ROBOTICS USE CASES
SEARCH AND RESCUE

Navigate unknown disaster environments

AGRICULTURE

Scout new field areas before mapping

DEFENSE

Autonomous reconnaissance in unmapped terrain

HARDWARE REQUIREMENTS
MIN VRAM
6GB
REC VRAM
12GB
MIN RAM
16GB
DISK SPACE
2GB
JETSON
Compatible
CPU ONLY
GPU Required
TESTED GPUs
RTX 3070RTX 4070Jetson AGX Orin
DEPLOYMENT PATHS
Local PC
SUPPORTED
Docker
SUPPORTED
ROS2
SUPPORTED
Jetson AGX Orin
SUPPORTED
Cloud
SUPPORTED
Isaac Sim
NOT SUPPORTED
RELATED PAPERS
NoMaD: Goal Masking Diffusion Policies for Navigation and Exploration
READ ON ARXIV →
RELATED ORGANIZATIONS
UC Berkeley
ORGANIZATIONS
UC Berkeley
DIFFICULTY
INTERMEDIATE

Requires some ML and Python experience

COMPUTE6/10
SETUP5/10
BEGINNER-FRIENDLY6/10
EDGE DEPLOYMENT
Jetson Compatible
CPU-only Not Available
Community →

Seeded and planned prompts — not live forum activity yet.