OPEN MODELS/NAVIGATION/GNM
BEGINNERNAVIGATION

GNM

The original General Navigation Model for cross-robot transfer

WHAT DOES THIS ACTUALLY DO?

GNM was one of the first navigation models to prove that a single trained model can drive completely different robot bodies to the same goal. It opened the door to everything that came after — including ViNT and NoMaD.

PARAMETERS
~15M
RELEASED
2023-01
LICENSE
MIT
ARCHITECTURE
ResNet...
CAN I ACTUALLY RUN THIS?
9
Beginner Friendly
OUT OF 10

Tiny, fast, and runs anywhere. If you're building your first mobile robot and want a navigation model, start here. It even runs on a Raspberry Pi.

DOWNLOAD WEIGHTS
View on GitHub
500MB disk space required
OVERVIEW

Goal-conditioned visual navigation policy trained on diverse cross-embodiment data. The predecessor to ViNT — controls various robots zero-shot and enables efficient fine-tuning.

TRAINED ON

Cross-embodiment navigation dataset (multiple robot platforms)

ARCHITECTURE
ResNet + Goal Encoder + MLP
CAPABILITIES
Goal-image conditioning
Cross-robot generalization
Topological map planning
Real-time inference
REAL ROBOTICS USE CASES
EDUCATION

Starter model for a mobile robot project

RESEARCH

Historical baseline for navigation benchmarks

WAREHOUSE

Basic goal-image navigation in structured environments

HARDWARE REQUIREMENTS
MIN VRAM
2GB
REC VRAM
4GB
MIN RAM
4GB
DISK SPACE
500MB
JETSON
Compatible
CPU ONLY
Supported
TESTED GPUs
Any GPUJetson NanoRaspberry Pi 5 (CPU)
DEPLOYMENT PATHS
Local PC
Extremely lightweightSUPPORTED
Docker
SUPPORTED
ROS2
SUPPORTED
Jetson Nano
Runs well on Jetson NanoSUPPORTED
Raspberry Pi 5
CPU mode worksSUPPORTED
Cloud
SUPPORTED
RELATED PAPERS
GNM: A General Navigation Model to Drive Any Robot
READ ON ARXIV →
RELATED ORGANIZATIONS
UC Berkeley
ORGANIZATIONS
UC BerkeleyResearch Community
DIFFICULTY
BEGINNER

Anyone with a computer can run this

COMPUTE9/10
SETUP7/10
BEGINNER-FRIENDLY9/10
EDGE DEPLOYMENT
Jetson Compatible
CPU-only Supported
Community →

Seeded and planned prompts — not live forum activity yet.