P³-WAM
A unified world-action model that adopts a shared video diffusion Transformer backbone with modality-specific adaptation layers, integrating perception, prediction, and planning capabilities under a single set of weights.
Timeline (last 90 days)
P³-WAM is a unified architecture containing three complementary modes—Perceive, Predict and Plan—that share the same video diffusion Transformer backbone.
P³-WAM's Perceive mode reconstructs articulated hand motion and camera trajectories from video.
P³-WAM's Predict mode generates future visual frames given actions.
Experiments show that unified training outperforms single-mode training on all three capabilities, including improved planning quality under vision-language model evaluation.
On four datasets, P³-WAM surpasses specialized baseline models on both 3D hand reconstruction and action-conditioned video prediction.
The goal-directed manipulation videos generated by P³-WAM are stronger than video generation baselines in instruction following and action progression.
The P³-WAM paper project page has been made public: https://p3-wam.pages.dev/.
All Facts (7)
P³-WAM is a unified architecture containing three complementary modes—Perceive, Predict and Plan—that share the same video diffusion Transformer backbone.
50%UnverifiedP³-WAM's Perceive mode reconstructs articulated hand motion and camera trajectories from video.
50%UnverifiedP³-WAM's Predict mode generates future visual frames given actions.
50%UnverifiedExperiments show that unified training outperforms single-mode training on all three capabilities, including improved planning quality under vision-language model evaluation.
50%UnverifiedOn four datasets, P³-WAM surpasses specialized baseline models on both 3D hand reconstruction and action-conditioned video prediction.
50%UnverifiedThe goal-directed manipulation videos generated by P³-WAM are stronger than video generation baselines in instruction following and action progression.
50%UnverifiedThe P³-WAM paper project page has been made public: https://p3-wam.pages.dev/.
50%