July 5, 2026

X-Humanoid unveils TG-VLA, the first whole-body Vision-Language-Action framework for humanoid robots

X-Humanoid has launched its TG-VLA, claiming it to be the world’s first full-size, whole-body Vision-Language-Action framework designed for humanoid robots. This innovative system addresses the limitations of current humanoid robots, which typically operate with disjointed actions, by enhancing the coordination of the robot’s entire body, including arms, legs, and torso. The TG-VLA framework integrates three core components: HEX for cross-embodiment learning, HAF-VLA to simplify complex motion into manageable steps, and DSRL-DCT for safe reinforcement learning in high-degree-of-freedom robots, enabling a 100% success rate in mobile manipulation tasks. This development is part of a growing trend in robotics focused on integrated control of vision, balance, memory, and multi-limb motion for more realistic and effective task performance.

HEX: HEX is a core component of TG-VLA for cross-embodiment learning across different humanoid robot bodies. It incorporates visual retrospection so robots can reference earlier visual information rather than treating each moment as isolated. This supports longer, context-dependent tasks by providing continuity beyond single camera frames.
TG-VLA: TG-VLA is the world’s first full-size whole-body Vision-Language-Action framework designed for humanoid robots. It integrates HEX, HAF-VLA, and DSRL-DCT to allow robots to understand tasks, remember scenes, predict outcomes, coordinate the full body, and adjust actions in real time. The framework emphasizes embodiment state prediction and whole-body manipulation over isolated limb control.
HAF-VLA: HAF-VLA is a TG-VLA component that decomposes whole-body actions into simpler structured steps. It refines motion through step-by-step denoising within a shared action space. The approach reduces the complexity of generating coordinated full-body movements.
DSRL-DCT: DSRL-DCT is the reinforcement learning module in TG-VLA for high-degree-of-freedom robots. It freezes the pretrained model, learns a compact noise-selection policy, and compresses the action space using Discrete Cosine Transform. This enables safer online learning while maintaining performance in mobile manipulation tasks.
X-Humanoid: X-Humanoid is a company developing advanced humanoid robotics systems. It unveiled TG-VLA as a full-size whole-body Vision-Language-Action framework to enable coordinated control across the entire robot body. The announcement highlights the company’s focus on addressing limitations in current humanoid robots for tasks requiring integrated balance, memory, and motion.

Whole-Body Control: Humanoid robot development increasingly targets integrated coordination of vision, balance, memory, and multi-limb motion for realistic task performance.
Embodied AI Frameworks: Vision-Language-Action models are being advanced to help robots maintain scene context and predict body states across extended sequences rather than reacting frame-by-frame.
Robotics Learning Techniques: Methods that break complex actions into hierarchical steps and compress high-dimensional spaces are gaining attention for safer training of multi-joint robotic systems.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like