BREAKING
Stanford open-sources GRaD-Nav
3DGS meets differentiable RL
3DGS
●
High-fidelity scenes
●
Differentiable representation
DiffRL
●
Gradients flow through sim
●
Faster policy learning
0
%
more efficient
0
%
of training time
GRaD-Nav++ success rates
Sim trained
83
Sim unseen
75
HW trained
67
HW unseen
50
VLA pipeline runs fully onboard
1
Vision + language in
↓
2
MoE routing
↓
3
Actions out
↓
4
Onboard flight
A reproducible drone baseline
AI NEWS BLITZ
A Stanford team has released open-source code for vision-based drone navigation using 3D Gaussian Splatting.