AI News AI资讯 5h ago Updated 1h ago 更新于 1小时前 50

Ant Group’s Robbyant Unveils LingBot-VA 2.0: A Causal Video-Action Model Built Natively for Physical AI 蚂蚁集团Robbyant发布LingBot-VA 2.0:专为物理AI构建的因果视频动作模型

Ant Group’s Robbyant introduces LingBot-VA 2.0, the first embodied-native foundation model that pretrains a causal video-action stack rather than fine-tuning existing video generators. The architecture replaces standard VAEs with a semantic visual-action tokenizer and employs a causal DiT with a sparse Mixture-of-Experts (MoE) video stream to address physical structure and latency limitations. Training utilizes multi-chunk prediction (MCP) and co-trains five diverse objectives (T2I, T2V, etc.) t 蚂蚁集团Robbinant发布LingBot-VA 2.0,首个原生具身基础模型,采用因果视频-动作架构而非微调视频生成器。 引入语义视觉-动作Tokenizers和因果DiT结合稀疏MoE的视频流,解决传统模型物理结构缺失和控制延迟问题。 通过多块预测(MCP)、五目标联合训练及前瞻推理(Foresight Reasoning)异步机制,显著提升训练效率与实时控制能力。 在RoboTwin 2.0基准测试中取得93.6%的平均成功率,结合一致性蒸馏和FP8优化,推理速度提升至225Hz。

72
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Ant Group’s Robbyant introduces LingBot-VA 2.0, the first embodied-native foundation model that pretrains a causal video-action stack rather than fine-tuning existing video generators.
  • The architecture replaces standard VAEs with a semantic visual-action tokenizer and employs a causal DiT with a sparse Mixture-of-Experts (MoE) video stream to address physical structure and latency limitations.
  • Training utilizes multi-chunk prediction (MCP) and co-trains five diverse objectives (T2I, T2V, etc.) to prevent catastrophic forgetting and accelerate convergence by 2.3x.
  • Deployment features a hierarchical VLM planner and an asynchronous "Foresight Reasoning" mechanism that overlaps execution with prediction, correcting drift via real-world observations.
  • LingBot-VA 2.0 achieves state-of-the-art performance on RoboTwin 2.0 (93.6% average success rate) and reduces inference latency from 927ms to 142ms per chunk through distillation and runtime optimizations.

Why It Matters

This release marks a significant shift from adapting digital content creation models to robotics toward building architectures natively designed for physical control. By addressing the structural mismatch between bidirectional video generation and causal robotic action, it offers a scalable path for generalist robot manipulation. The integration of asynchronous foresight reasoning provides a practical solution to the latency bottlenecks that typically hinder real-time embodied AI applications.

Technical Details

  • Native Embodied Architecture: Unlike prior works that fine-tune video diffusion models, LingBot-VA 2.0 pretrains a causal Diffusion Transformer (DiT) natively for embodiment, ensuring temporal consistency aligns with forward-in-time control.
  • Semantic Visual-Action Tokenizer: Replaces reconstruction-only VAEs with a tokenizer that aligns visual latents with a perception encoder and extracts compact transition variables, allowing unlabeled web video to provide action-relevant supervision.
  • Sparse MoE Video Stream: The video expert utilizes a sparse MoE layer with 128 SwiGLU experts and top-8 routing, while the action expert remains dense. This asymmetric scaling results in ~13B total parameters with only ~2.5B active per token at inference.
  • Multi-Chunk Prediction (MCP): To fix myopic supervision, MCP attaches lightweight modules predicting the next three chunks, enabling the model to learn long-term dependencies and significantly speeding up training convergence.
  • Asynchronous Foresight Reasoning: The system runs prediction and execution in parallel streams. The video expert imagines outcomes while the robot acts, and real observations overwrite imagined latents to correct drift, supported by a VLM planner emitting structured JSON instructions.

Industry Insight

  • Shift to Native Pretraining: Developers should move away from retrofitting video generation models for robotics. Building causal, physically grounded stacks from scratch yields better alignment with control requirements and reduces the need for extensive domain adaptation.
  • Latency Management via Asynchrony: For real-time deployment, decoupling planning/prediction from execution is critical. Implementing asynchronous buffers and foresight mechanisms allows robots to maintain high-frequency control loops despite heavy model inference times.
  • Efficiency Through Sparsity and Distillation: Combining sparse MoE architectures with consistency distillation and low-precision execution (FP8) is essential for making large-scale embodied models viable on edge hardware. These techniques reduce inference time by over 6x without significant accuracy loss.

TL;DR

  • 蚂蚁集团Robbinant发布LingBot-VA 2.0,首个原生具身基础模型,采用因果视频-动作架构而非微调视频生成器。
  • 引入语义视觉-动作Tokenizers和因果DiT结合稀疏MoE的视频流,解决传统模型物理结构缺失和控制延迟问题。
  • 通过多块预测(MCP)、五目标联合训练及前瞻推理(Foresight Reasoning)异步机制,显著提升训练效率与实时控制能力。
  • 在RoboTwin 2.0基准测试中取得93.6%的平均成功率,结合一致性蒸馏和FP8优化,推理速度提升至225Hz。

为什么值得看

本文揭示了具身智能从“数字内容生成”向“物理世界因果控制”转型的关键技术路径,为机器人通用操作提供了新的基础模型范式。其异步前瞻推理和混合专家架构设计,有效解决了长视距规划与低延迟执行之间的矛盾,对行业落地具有极高的参考价值。

技术解析

  • 原生因果架构:摒弃了传统的双向注意力机制,预训练原生的因果DiT(Diffusion Transformer)。视频专家采用稀疏MoE路由层(128个SwiGLU专家,Top-8路由),而动作专家保持密集FFN,实现非对称扩展,总参数量约15.3B,推理激活约2.5B。
  • 语义视觉-动作Tokenizers:第一阶段替换仅用于重建的VAE,引入感知编码器教师进行语义对齐,并提取潜在动作变量。世界状态和动作共享同一潜在空间,使得无标签网络视频也能提供动作相关的监督信号。
  • 多块预测(MCP)与联合训练:为解决教师强制导致的短视问题,MCP模块预测后续三个块,使训练速度提升2.3倍。同时,文本到图像(T2I)、文本到视频(T2V)等五个目标联合训练,避免遗忘早期先验。
  • 前瞻推理与异步执行:部署时采用异步流,机器人在执行当前动作块时,视频专家并行想象下一状态结果,动作专家据此解码下一步动作。真实观测返回后覆盖想象状态,并通过前向动力学接地损失校正漂移。
  • 高性能加速策略:通过一致性蒸馏将视频和动作采样步数分别降至2步,结合FP8 TensorRT引擎、Paged/Ragged KV缓存及FlashInfer注意力优化,最终实现142ms/chunk的推理延迟和225Hz的控制频率。

行业启示

  • 具身模型需回归物理因果性:通用视频生成模型直接应用于机器人存在物理结构理解不足和控制延迟高的缺陷,原生构建因果视频-动作模型是提升机器人泛化能力的必由之路。
  • 异步并行是突破实时性瓶颈的关键:通过前瞻推理将预测与执行解耦,利用计算冗余掩盖模型延迟,是实现高频闭环控制的有效工程手段,尤其适用于对响应时间敏感的硬件平台。
  • 数据利用与训练效率并重:利用无标签网络视频进行语义对齐,结合多块预测和混合优化器,证明了在有限标注数据下通过架构创新和训练策略优化可大幅提升模型性能与收敛速度。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Robotics 机器人 Video Generation 视频生成 Multimodal 多模态 Research 科学研究 Product Launch 产品发布