Baseline→optimized policy across procedurally generated wildfire terrain
How it works
Three layers, one rollout.
01 · Data
Isaac Sim
Trajectories generated in NVIDIA Isaac Sim across procedurally varied terrain. Slopes, gravel, mixed friction, randomized payloads, all recorded into the training set.
02 · Model
CNN world model
Trained on a CNN architecture that predicts the next state from the current observation. The policy commits to a step only when the world model agrees it stays upright.
03 · Validate
Custom evals
Every checkpoint runs a held-out suite of incline, payload, and friction combinations. Pass rate gates promotion to the rollout viewer.
Try it
Reduced fall rate by 70.59%
From baseline to optimized policy across procedurally generated wildfire terrain.