KDE
VAE
Normalizing flow
Diffusion
NeuralODE
MOBILE
MBPO
GORMPO
Penalize rollouts that leave the data’s support
+
Click to show details
Generative OOD-regularization
Model-based policy
rollouts
penalized reward
Generative OOD-regularization
Model-based policy
4. Density-based guardian
likelihood of next state and action
Penalized
reward
1. Replay buffer
real data
3. Dynamics
model ensemble
5. Replay buffer
generated data
predict next state, reward
penalize
store
sample
2. Policy
optimization
action
sample
sample
sample
2. Policy
optimization
Model–Bellman
inconsistency
sample
action
ensemble targets
penalize
critic
sample
Sparse D4RL
Medium-expert, 22–28% of an unsafe region removed
density
action
1
0.1
0.01
0.001
reward
reward
reward
halfcheetah
hopper
walker2d
Our medical dataset
Cardiogenic shock, 379 patients on MCS
Impella
sparse reward–action
Policy evaluation
MuJoCo tasks
+ click a dataset to see results
MOBILE
GORMPO
return gain
halfcheetah
hopper
walker2d
Impella weaning
+ click a dataset to see results
✓
Blood pressure stays above 60: the patient is stable, so the pump should be
weaned
.
CORMPO
no MOBILE-based
medical policy
GORMPO
blood pressure
observed
P-level
clinician decision
critical threshold