WAM-TTT: A New Robot Post-Training Framework

新智元 · wechat · 2026-07-16

The article introduces WAM-TTT by Galbot, a test-time post-training framework for embodied intelligence models designed to improve robot generalization in real-world deployment environments.

Core Concept

Instead of relying on expensive robot teleoperation trajectories during deployment, it directly uses egocentric (first-person) human videos.
The input can be pure video without action annotations, relying on self-supervised video prediction to learn task dynamics.
It employs a "frozen backbone + lightweight fast weight update" approach to preserve pre-training capabilities and minimize catastrophic forgetting.

Results & Comparisons

Cross-environment testing shows that WAM-TTT significantly outperforms baselines when facing changes in lighting, objects, and backgrounds.
It maintains a much higher generalization rate compared to ICL baselines that require human video context.
Under the same budget, the authors' ablation results show that 100 robot trajectories + 100 human videos perform almost on par with 200 robot trajectories.

Key Argument

The article positions this work as a crucial step toward the scaled deployment of embodied AI: after learning skills in training environments, robots can adapt to new scenarios at a fraction of the data, time, and financial cost. The author also emphasizes that Chinese teams are beginning to establish their own technological paradigms in embodied intelligence methodologies.

Original post →

WAM-TTT: A New Robot Post-Training Framework

More from Embodied