TARS unveils AWE 3.5, an embodied-native model with 'Born as One' architecture

rohanpaul_ai · x · 2026-08-22

TARS has introduced the AI World Engine (AWE) 3.5, an embodied-native foundation model designed to address the issue of current embodied AI often feeling like bolted-on modules.

Its core "Born as One" architecture integrates action, perception, geometry, and touch into a single model from the start, rather than stitching them together later. This model-driven system aims to generalize across different tasks, objects, environments, and robot bodies. The training methodology implements a full closed-loop process: pre-training gives the model a base understanding of action patterns, spatial structure, and physical laws via two priors, while post-training uses the AI World Engine to roll out possible future states for validation.

Original post →

More from Embodied

Embodied channel →