← All models

Hy-Embodied-0.5-VLA-RoboTwin

Published Publisher Tencent

Overview

Hy-Embodied-0.5-VLA-RoboTwin is a vision-language-action (VLA) checkpoint from Tencent Robotics X and the Tencent Hy Team. It is the supervised fine-tuned RoboTwin variant of Hy-Embodied-0.5-VLA, trained on all 50 bimanual manipulation tasks in the RoboTwin 2.0 benchmark. Tencent announced its release on June 15, 2026. (huggingface.co)

Capabilities and intended use

  • Processes visual observations and language instructions to generate continuous robot action chunks for manipulation. Its architecture combines the Hy-Embodied-0.5 vision-language backbone, a flow-matching action expert and a compact temporal memory encoder. (github.com)
  • This checkpoint uses six frames—one current and five historical—and an action horizon of 20 at a three-times-downsampled rate. (huggingface.co)
  • Intended for robot-learning research, RoboTwin 2.0 evaluation and development of bimanual manipulation policies. Tencent reports average success rates of 90.9% on Clean and 90.1% on Randomized RoboTwin 2.0 settings, averaged across 50 tasks and 100 rollouts per task. These are benchmark results, not guarantees for real-world deployment. (huggingface.co)

Training and availability

The checkpoint was initialized from the UMI-pretrained Hy-VLA model and fine-tuned on RoboTwin 2.0 demonstrations. Tencent reports that the broader pretraining corpus contains over 10,000 hours of UMI demonstrations, while over 2,000 hours of data were released publicly. The checkpoint and code are available for self-hosted use; the model weights are licensed under Apache 2.0. (huggingface.co)

Limitations

This is a RoboTwin-specialized checkpoint, not a general-purpose conversational model. Results on the simulated benchmark should not be assumed to transfer unchanged to a different robot, environment or task. Deployment on physical hardware requires appropriate integration, validation and safety precautions. Tencent's released dataset is only a portion of the full pretraining corpus. (huggingface.co)

Pricing

As of November 16, 2026, model-specific official API pricing could not be verified in Tencent Cloud's mainland China or international pricing materials. Input and output charges are unverified; no CNY or USD amount is asserted. The availability of downloadable weights does not establish that hosted inference, computing or deployment is free. (cloud.tencent.com)

Official sources