Tencent Robotics X and Foton Lab Release HyVLA-0.5: Open-Source VLA Model Achieves SOTA on RoboTwin 2.0
Tencent Robotics X, in collaboration with Foton Lab and the Hunyuan team, has released HyVLA-0.5 (Hy-Embodied-0.5-VLA) — an end-to-end embodied intelligence model designed for real-world robotic manipulation tasks.
Key Technical Innovations
Self-Developed UMI Data Collection System
HyVLA-0.5 is built on a proprietary sub-millimeter precision finger-sleeve UMI (Universal Manipulation Interface) data collection hardware-software system (patent: 2025020117CN). This system enables:
- 10,000+ hours of human demonstration data collection
- No teleoperation data required during supervised training phase
- High deployment success rates across multiple robot embodiments
Benchmark Performance
On the RoboTwin 2.0 simulation benchmark, HyVLA-0.5 achieves over 90% success rate in both simple and complex settings — making it the current State-of-the-Art (SOTA) open-source VLA model on the leaderboard.
Proximalized Preference Optimization (PRO)
HyVLA-0.5 is the first model to systematically introduce Proximalized Preference Optimization (PRO) into flow-matching-based VLA reinforcement post-training. By fully leveraging real robot failure data during training, the model achieves:
- Near-100% success rate on real-world robotic tasks
Significance
The release represents a meaningful step forward in open-source embodied AI, demonstrating that:
- High-quality human demonstration data collected via precision wearable devices can effectively train generalist robot manipulation policies
- Systematic use of failure data through preference optimization can dramatically improve real-world deployment reliability
- The combination of large-scale demonstration data and reinforcement post-training is a viable path toward production-ready embodied AI systems
The model and technical details are available through Tencent's research channels.
This article is based on the official announcement from Tencent Robotics X, Foton Lab, and the Hunyuan team.
Embodied Global tracks 1,000+ articles on 53 companies across the embodied AI ecosystem. embodiedglobal.com
