P3: Probabilistic Policy Propagation for Stable VAE-Based Robot Learning

Researchers introduced P3, a distribution-aware method that makes VAE-based PPO training more stable and efficient in robot learning.

P3 replaces single-sample latent estimates with probabilistic propagation and calibration for VAE-based robot policies.

The approach improves training stability, sample efficiency, and convergence for stochastic latent policy optimization.

It also delivers better performance on challenging humanoid parkour tasks.