TienKung Robotics launches TG-VLA for full-body humanoid control
The X-Humanoid team is trying to move humanoids from arm demos to coordinated whole-body work.
By Ryan Merket ยท Published
Why it matters
TG-VLA shows how humanoid competition is shifting from walk-and-wave demos toward the harder software layer that coordinates perception, language, balance, and manipulation.
TienKung Robotics has launched TG-VLA, a full-size, full-body vision-language-action framework for humanoid robots, according to posts from the X-Humanoid account on X and reporting from Shanghai Securities News via Sina Finance.
The Beijing Humanoid Robot Innovation Center, which develops the TienKung line, positions TG-VLA as a stack that connects three layers of humanoid autonomy: high-level semantic decision-making, whole-body coordinated motion, and adaptation in dynamic environments.
The center also describes TG-VLA as the world's first full-size, full-body VLA framework. The broader VLA category is already active. Google DeepMind's RT-2 showed in 2023 that vision-language models could be adapted into robot policies that directly output actions. OpenVLA, released in 2024, pushed open-source VLA models for robotic manipulation. Recent research projects such as WholeBodyVLA and WOLF-VLA have also attacked whole-body humanoid loco-manipulation and locomotion. TienKung Robotics' narrower assertion is about a full-size humanoid framework that spans the whole body, rather than a claim that it invented VLA robotics.
Why the full-body framing matters: humanoids often fail in the coupling between cognition and mechanics. A model can infer that a robot should squat to pick something up, but the body still has to maintain contact, distribute force, plan a reachable trajectory, and recover when real-world friction or object placement differs from training data. VLA research is moving toward that coupling, but production-grade humanoid work also has to contend with latency, safety margins, battery limits, actuator wear, and the data cost of collecting reliable full-body demonstrations.
For developers watching this space, the center's Open X-Humanoid GitHub organization hosts code and simulation assets tied to the TienKung platform. The next proof point for TG-VLA will be reproducible evaluations and side-by-side tasks that show coordinated whole-body behavior beyond curated demos.