Tencent Robotics XChina2026

Hy-Embodied-0.5-VLA

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

RobotsS1by Astribot CG1by Unitree Robotics C

Key figures as stated by the developer

5 of 6 stated · 1 gap

Model type ⁠A vision-language-action
Controls ⁠A arms and hands
Inputs ⁠A camera language robot state
Access ⁠A open download
Parameters not stated
Base model ⁠A Hy-Embodied-0.5 MoT

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (6)

Model type

Developer wording: "an end-to-end Vision-Language-Action system".

Controls

"Actions are predicted as relative-to-current-frame end-effector delta chunks". The released data and checkpoints are for two-armed manipulation.

Inputs

The quick start feeds camera views, a task text such as "pick up the bottle" and a robot state described as "Normalized dual-arm EEF".

Access

Tencent released the model, "including the codebase", the checkpoints and the data.

Parameters

Note No total for the whole model; the only count given is for one part: "a dedicated 370M-parameter action expert".

Base model

The README lists "Backbone — Hy-Embodied-0.5 MoT" as the backbone.

Full specifications

9 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action`	A Developer wording: "an end-to-end Vision-Language-Action system".
Inputs	`camera` `language` `robot state`	A The quick start feeds camera views, a task text such as "pick up the bottle" and a robot state described as "Normalized dual-arm EEF".
Controls	`arms and hands`	A "Actions are predicted as relative-to-current-frame end-effector delta chunks". The released data and checkpoints are for two-armed manipulation.
Degrees of freedom controlled	not stated	Note The actions are end-effector poses per arm, not joint commands: "10-dim per arm: xyz + rot6d + gripper". ⁠B
Size and speed
Parameters	not stated	Note No total for the whole model; the only count given is for one part: "a dedicated 370M-parameter action expert". ⁠A
Control rate	50Hz	C Action execution layer; the technical report says: "For real-world deployment, actions are sampled at 50 Hz". The released pre-trained checkpoint predicts actions "H=50 at 10 Hz".
Where it runs and what it is built on
Runs where	not stated	Note Tencent lists "(≥ 16 GB VRAM recommended)" but does not say whether the model runs on the robot. ⁠A
Base model	Hy-Embodied-0.5 MoT	A The README lists "Backbone — Hy-Embodied-0.5 MoT" as the backbone.
Openness
Access	open download	A Tencent released the model, "including the codebase", the checkpoints and the data.
License	Apache-2.0	B Model weights: "This model is released under" Apache-2.0. Code: "Hy-Embodied-0.5-VLA is licensed under the Apache-2.0."
The developer’s own claims
Learns new tasks	needs retraining	A Developer claim: the pre-trained checkpoint is "intended as a generalist starting point for fine-tuning".
Works across robots	yes	A Developer claim: "demonstrates robust cross-embodiment transfer across four real-world robot platforms". The technical report names "Track B — Cross-Embodiment Transfer (JAKA K1, Astribot S1)". Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

Tencent Robotics X describes Hy-Embodied-0.5-VLA (short name Hy-VLA) as an end-to-end vision-language-action system built on the Hy-Embodied-0.5 backbone.
Two checkpoints are released: a pre-trained generalist one (Hy-VLA-UMI) and one post-trained on a simulation benchmark (Hy-VLA-RoboTwin).
Tencent's HY-Embodied repository also calls the release HY-VLA-0.5.
Tencent released the code, the checkpoints and more than 2,000 hours of two-armed demonstration data, recorded with a fingertip capture rig, on 15 June 2026.

Sources

A Manufacturer product page. github.com/Tencent-Hunyuan/Hy-Embodied-0.5-VLA · Retrieved 2026-10-13
B huggingface.co/tencent/Hy-Embodied-0.5-VLA-UMI · Retrieved 2026-10-13
C arxiv.org/html/2606.14409 · Retrieved 2026-10-13

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.