Breaking

Stanford Releases GRaD-Nav, a 3DGS-Based Visual Drone Navigation Framework

June 29, 2026 at 13:48 EDT

Robotics
Research & Papers
Open Source

A research team at Stanford University has released the papers and code for GRaD-Nav, a visual drone navigation framework that combines 3D Gaussian Splatting (3DGS) with differentiable dynamics. The work, led by Qianzhong Chen and colleagues, integrates photorealistic scene representation with efficient gradient-based learning, and is notable for transferring simulation-trained policies to real hardware without additional fine-tuning. The official GitHub repository provides training and testing scripts along with datasets.

Stanford Research · Open-Source Release

GRaD-Nav: vision-based drones that learn to fly faster

Code for two papers (IROS 2025 & RA-L 2025) is now public — fusing 3D Gaussian Splatting with differentiable reinforcement learning so gradients flow through simulation, training policies far more efficiently and transferring to real drones with zero fine-tuning.

+300%

Sample-efficiency gain vs conventional RL

20%

Of the training time to reach a better policy

Zero

Shot sim-to-real transfer, no fine-tuning

Training time to a better policy

Same goal — GRaD-Nav reaches it in a fraction of the effort

100%

Conventional RL

20%

GRaD-Nav

GRaD-Nav++ success rates

Trained tasks vs unseen tasks — in simulation and on real hardware

Simulation

83%

Trained

75%

Unseen

Hardware

67%

Trained

50%

Unseen

Multi-environment adaptation (average): Sim 81% · Real 67%

How it learns

3D Gaussian Splatting

High-fidelity, differentiable scene

→

Differentiable RL

Gradients flow through sim

→

Onboard Policy

Flies real drones, no infrastructure

GRaD-Nav++ — Vision-Language-Action framework

Outputs actions directly from visual and language inputs
Mixture-of-Experts routing for multi-task generalization, less forgetting
Runs fully onboard, independent of external infrastructure

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.