Galaxy General Launches World's First Test-Time Training Framework for Embodied AI Robots, WAM-TTT

Galaxy General today announced WAM-TTT (World-Action Model Test-Time Training), the world's first test-time training framework designed for embodied intelligence large models, marking the first successful transfer of the TTT paradigm from large language models to physical robot control.

This breakthrough means robots can continuously learn and adapt during deployment rather than relying solely on pre-training knowledge. Until now, the entire embodied AI industry has been stuck with a core problem: a robot that masters box-stacking in a training lab fails immediately when faced with a different environment or different boxes.

WAM-TTT's key innovation is that it requires no additional robot trajectory data collection or human action annotation during deployment. Instead, it adapts using only unlabeled human RGB video footage of the target scene. The framework is built on a pre-trained World Action Model (WAM) with two components — a video expert and an action expert — connected through joint attention.

Crucially, WAM's main weights remain completely frozen throughout the process. All learning occurs in a lightweight fast-weight memory module. This is analogous to a chef jotting down a customer's request on a sticky note, then executing the dish using existing culinary skills guided by that note.

In experiments, WAM-TTT achieved a 74.1% average task success rate when trained on 100 robot trajectories plus 100 human videos — comparable to training on robot trajectories alone. Under certain conditions, a single human video can nearly 1:1 replace an expensive robotic teleoperation data sample.

By contrast, traditional approaches using human pose estimation and action retargeting achieved only 28.9% average completion across four tasks — 43.4 percentage points below WAM-TTT. In unfamiliar real home environments, WAM-TTT maintained approximately 75.6% capability retention, while in-context learning baselines collapsed from 48.4% to 7.1%.

This framework is technically far more challenging than TTT in the NLP domain. Robot action spaces are high-dimensional and continuous, where a millimeter-level grip angle error can mean the difference between success and failure, and physical interactions are irreversible once executed. WAM-TTT offers a path forward that diverges from the industry's default strategy of collecting ever-larger robot datasets.