Hy-Embodied-0.5-VLA：从视觉-语言-动作模型到真实世界机器人学习栈

Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack

孙兴武 Xingwu Sun · Tencent Hunyuan · 2026-06-12 · arXiv:2606.14409 ↗

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

在本报告中，我们介绍了 Hy-Embodied-0.5-VLA（简称 HyVLA-0.5），这是一个端到端系统，涵盖了完整的机器人学习栈：数据收集、模型设计、持续预训练和监督微调、强化学习后训练以及真实世界部署。每个组件在这个栈中都扮演着独特的角色。

In this report, we present Hy-Embodied-0.5-VLA, abbreviated as HyVLA-0.5, an end-to-end system that spans the full robot learning stack: data collection, model design, continued pre-training and supervised fine-tuning, RL post-training, and real-world deployment. Each component serves a distinct role in this stack.

核心贡献 · Key contributions

• 端到端机器人学习栈，涵盖数据收集、模型设计、预训练、微调、强化学习后训练和部署。
  End-to-end robot learning stack spanning data collection, model design, pre-training, SFT, RL post-training, and deployment.
• 定制指尖 UMI 设备与动作捕捉系统，实现亚毫米精度的人类演示。
  Custom fingertip UMI device with motion-capture cage for sub-millimeter-precision human demonstrations.
• 具身原生混合 Transformer 骨干网络与流匹配动作专家，实现连续高频控制。
  Embodied-native MoT backbone with flow-matching action expert for continuous, high-frequency control.
• 增量块动作表示，将策略与具身特定运动学解耦。
  Delta-chunk action representation decoupling policy from embodiment-specific kinematics.
• FlowPRO：无评论家、无奖励的离线强化学习，利用偏好优化与对比梯度取消。
  FlowPRO: critic-free, reward-free offline RL using preference optimization with contrastive gradient cancellation.
• 异步推理框架与三次贝塞尔动作平滑器，实现 C1 连续部署。
  Asynchronous inference framework with cubic Bézier action smoother for C1-continuous deployment.

局限 · Limitations

• UMI 数据依赖动作捕捉系统，限制了野外部署的灵活性。
  UMI data collected with motion-capture cage limits in-the-wild deployment flexibility.
• 跨具身迁移仅在有限任务和机器人上评估。
  Cross-embodiment transfer evaluated only on a limited set of tasks and robots.
• FlowPRO 需要遥操作干预与回滚流程来收集偏好数据。
  FlowPRO requires teleoperated intervention-and-rollback pipeline for preference data collection.
• 部署时的执行速度与安全性权衡尚未完全解决。
  Deployment-time execution speed and safety trade-offs not fully addressed.
• 未研究零样本泛化；数据规模可能仍不足。
  Zero-shot generalization not studied; data scale may still be insufficient.

论文章节 · Sections（共 23）

• 1 引言 1 Introduction
• 2 模型架构 2 Model Architecture
• 2.1 问题形式化 2.1 Problem Formulation
• 2.2 Hy-Embodied：模态自适应计算骨干 2.2 Hy-Embodied: Modality-Adaptive Computing Backbone
• 2.3 双塔流匹配动作专家 2.3 Action Expert with Dual-Tower Flow Matching
• 2.4 时空注意力紧凑记忆编码器 2.4 Compact Memory Encoder with Temporal-Spatial Attention
• 3 预训练与监督微调 3 Pre-training and Supervised Fine-tuning
• 3.1 Hy-UMI-10K：高保真操作数据集 3.1 Hy-UMI-10K: High-Fidelity Manipulation Dataset
• 3.2 预训练 3.2 Pre-training
• 3.3 监督微调 3.3 Supervised Fine-tuning
• 4 强化学习后训练 4 Reinforcement Learning Post-Training
• 4.1 设计原则 4.1 Design Principles
• 4.2 方法 4.2 Method
• 5 部署 5 Deployment
• 5.1 具身无关平台映射 5.1 Embodiment-Agnostic Platform Mapping
• 5.2 实时控制异步执行 5.2 Asynchronous Execution for Real-Time Control
• 5.3 延迟感知贝塞尔块拼接 5.3 Latency-Aware Bézier Chunk Stitching
• 6 评估 6 Evaluation
• 6.1 模拟任务 6.1 Simulated Tasks
• 6.2 真实世界任务 6.2 Real-World Tasks
• 6.3 真实世界强化 6.3 Real-World Reinforcement
• 7 相关工作 7 Related Work
• 8 讨论 8 Discussion

阅读逐段中英对照全文 →