G1 Garden Walk — Gaussian Splat Scene

Photorealistic digital twin of a real garden, built from multi-view photographs via 3D Gaussian Splatting (3DGRUT, PSNR 26.58 dB) with PPISP neural rendering. Unitree G1 humanoid deployed into the reconstructed scene running trained RL locomotion policy on NVIDIA Isaac Sim 6.0.1 with Newton physics solver.

Technical details:

Scene source: mipnerf360/garden dataset

Ground-slope adaptation: GROUND_SLOPE=0.09, tilt correction 1.820°

Physics: Newton solver, foot-contact validated on deformable grass/slope terrain

4 render cameras, shader cache warm 48s

85+ scripts, fact-tracking database (track.py)

Highlights:

Real-world scene reconstructed as interactive digital twin

G1 locomotion tested on uneven terrain with grass and slopes

Physics simulation validated foot-contact on deformable surfaces

Published as interactive showcase on forenly.ai

Source: github.com/ForenlyAI/g1-skill-layer-sim (private)