opentau.policies.pi05_mem

PI05 Mem Policy Module.

This module implements a variant of the π05 (Pi05) Vision-Language-Action Flow Model with the low-level memory architecture from Torne, Pertsch, Walke et al. “MEM: Multi-Scale Embodied Memory for Vision Language Action Models”. The SigLIP image encoder is extended with space-time separable attention every N-th layer so past frames can inform the current-frame tokens without introducing new learnable parameters, and temporal state sequences are projected into one continuous token per timestep for the Gemma backbone.

Modules

configuration_pi05	Configuration module for the PI05 Mem Policy.
modeling_pi05	π05 Mem: A Vision-Language-Action Flow Model with space-time SigLIP video encoding and temporal state sequences.
motion_module	Space-Time Self-Similarity (STSS) motion module.
rldx_video_encoder	RLDX SigLIP video encoder: plain SigLIP + STSS motion module.