X-Humanoid introduces TG-VLA with claimed 100% mobile-manipulation success

• According to rohanpaul_ai's launch thread, X-Humanoid calls TG-VLA the world’s first full-size, whole-body Vision-Language-Action framework for humanoid robots.
• The TG-VLA overview breaks the framework into three pieces: HEX for cross-embodiment learning, HAF-VLA for structured whole-body action, and DSRL-DCT for reinforcement learning on high-DoF robots.
• The DSRL-DCT note says X-Humanoid claims 100% success in mobile-manipulation tasks by freezing the pretrained VLA, learning a smaller noise-selection policy, and compressing the action space with Discrete Cosine Transform.
• The HEX follow-up describes visual retrospection, where the robot can reuse earlier visual context instead of treating every moment as a fresh frame.

The thread links out to the paper and project page. The HEX follow-up names visual retrospection for scene memory, the HAF-VLA note describes denoising inside one shared action space, and the DSRL-DCT note says the RL layer compresses a large robot action space with Discrete Cosine Transform.

X-Humanoid’s frame is whole-body coordination: task understanding, scene memory, prediction, full-body movement, and online adjustment during execution, according to the launch thread.

The concrete bet is that humanoids need a single coordinated action chain, not just better grasping. The hard case is a task where torso, legs, arms, hands, balance, vision, memory, and timing all interact.

TG-VLA’s stated chain:

1. Understand the task.
2. Remember the scene.
3. Predict what may happen next.
4. Coordinate the whole body.
5. Adjust while acting.

HEX is the cross-embodiment piece. The launch overview names it as the component for learning across different humanoid robot bodies, while the follow-up adds visual retrospection: the robot can look back at earlier visual context instead of treating each frame as a fresh scene.

That makes HEX the memory-ish layer in the stack. Longer manipulation tasks need context from before the current camera frame, especially when the body has moved and the scene has changed.

According to the HAF-VLA note, the system avoids generating every whole-body movement at once. It breaks motion into structured action steps, then cleans the final motion through denoising.

The mechanics are compact:

• Break whole-body action into simpler steps.
• Run step-by-step denoising.
• Keep the output inside one shared action space.

That shared-space detail is the interesting part. It suggests the model is not stitching separate arm, base, and torso controllers after the fact.

rohanpaul_ai's DSRL-DCT note attributes the 100% mobile-manipulation success claim to X-Humanoid’s reinforcement-learning layer.

The reported recipe has three parts:

• Freeze the pretrained Vision-Language-Action model.
• Learn a smaller noise-selection policy.
• Compress the robot’s large action space into a smaller latent space using Discrete Cosine Transform.

That is the engineering move worth remembering: the RL update targets a smaller control problem instead of retraining the full VLA against a high-degree-of-freedom robot.

The linked paper is titled “HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation,” according to the paper-link post. TG-VLA is the framework label in the launch thread; HEX is the named contribution in the paper title.

The same post also points to a project page, which is the most direct artifact in the evidence pool for readers who want the underlying demos and paper materials.