A Pragmatic VLA Foundation Model (LingBot-VLA)

01 动机Motivation

当前 VLA 模型面临两大核心挑战：数据规模不足与训练效率低下。大多数方法仅在数千小时数据上训练，且缺乏能在海量数据上高效扩展的训练框架，难以验证真正意义上的 scaling law。LingBot-VLA 的目标是构建一个务实可用的 VLA 基础模型——覆盖 9 种真实双臂机器人、2 万小时遥操作数据，并在 100 项任务的标准化基准上与主流方法直接对比。Current VLA models face two core challenges: insufficient data scale and low training efficiency. Most methods are trained on only a few thousand hours of data and lack a training framework that scales efficiently to massive data, so a genuine scaling law is hard to verify. LingBot-VLA aims to build a pragmatic VLA foundation model — covering 9 real dual-arm robots and 20000 hours of teleoperation data, and compared directly with mainstream methods on a standardized benchmark of 100 tasks.

"We present LingBot-VLA, a vision-language-action system trained on around 20,000 hours of real-world data from 9 popular dual-arm robot configurations."

LingBot-VLA overview — 图1：LingBot-VLA 总览。左侧展示数据采集所用的 9 种双臂机器人平台（共约 2 万小时），中间为 Mixture-of-Transformers 模型架构，右侧为跨平台的 GM-100 百任务实体评测结果。Figure 1: Overview of LingBot-VLA. Left: the 9 dual-arm robot platforms used for data collection (around 20000 hours in total); middle: the Mixture-of-Transformers model architecture; right: cross-platform real-robot evaluation on the GM-100 hundred-task benchmark.

~20,000h真实遥操作预训练数据real teleoperation pre-training data

9种双臂机器人平台dual-arm robot platforms

+4.28%平均 SR 超越 π₀.₅（GM-100）average SR over π₀.₅ (GM-100)

261samples/s（8-GPU 训练吞吐）samples/s (8-GPU training throughput)

02 方法Method

LingBot-VLA 以 Qwen2.5-VL 为视觉语言骨干，引入独立的"action expert"模块，通过 Mixture-of-Transformers (MoT) 框架将视觉语言与动作两条通路解耦，再经共享 self-attention 机制耦合；动作输出采用 Flow Matching 预测 50-token 动作块；深度信息通过 vision distillation 蒸馏融入，进一步提升空间感知能力。LingBot-VLA takes Qwen2.5-VL as its vision-language backbone and introduces a separate "action expert" module. A Mixture-of-Transformers (MoT) framework decouples the vision-language and action pathways, which are then coupled through a shared self-attention mechanism; the action output uses Flow Matching to predict 50-token action chunks; depth information is fused in through vision distillation, further improving spatial perception.

Pre-training dataset visualization — 图2：LingBot-VLA 预训练数据集可视化。9 种机器人平台的视觉观测示例，涵盖多样化的操作场景与任务类型；标注流程先由人工标注员将视频分解为原子动作，再由 Qwen3-VL-235B 自动生成语言指令。Figure 2: Visualization of the LingBot-VLA pre-training dataset. Visual observations from 9 robot platforms cover diverse manipulation scenes and task types; in the annotation pipeline, human annotators first decompose the videos into atomic actions, and Qwen3-VL-235B then generates the language instructions automatically.

Mixture-of-Transformers (MoT) 架构Mixture-of-Transformers (MoT) architecture

视觉语言路径与动作路径各自拥有独立的 Transformer 层（MLP、layer norm 等），通过 blockwise causal attention 实现模态间解耦：观测块与动作 token 分别处理，避免跨模态梯度干扰。共享的 self-attention 则保证两条路径能够互通信息，实现高效的跨模态对齐。输入观测包含三视角 RGB-D 图像、自然语言任务指令和机器人本体感知状态。The vision-language pathway and the action pathway each own separate Transformer layers (MLP, layer norm, etc.), and blockwise causal attention decouples the two modalities: observation blocks and action tokens are processed separately, avoiding cross-modal gradient interference. The shared self-attention still lets the two pathways exchange information, achieving efficient cross-modal alignment. The input observation contains three-view RGB-D images, a natural-language task instruction and the robot proprioceptive state.

Flow Matching 动作预测 + Vision DistillationFlow Matching action prediction + Vision Distillation

动作头采用 Flow Matching 对连续动作分布建模，每次预测 50-token 的动作序列块，支持快速推理。深度感知模块以可学习 query 与深度 token 对齐，并引入蒸馏损失（distillation loss）将专用深度模型的空间感知能力迁移至视觉骨干，无需推理时额外调用独立深度模型。训练效率优化包括：Fully Sharded Data Parallel (FSDP)、针对 action expert 的 shard 分组（减少通信开销）、bfloat16 存储 + float32 规约的混合精度策略、FlexAttention 稀疏多模态注意力，以及 torch.compile 算子融合——最终实现 GPU 数量从 8 到 256 的近线性吞吐扩展。The action head models the continuous action distribution with Flow Matching, predicting a 50-token action chunk at a time and supporting fast inference. The depth-perception module aligns learnable queries with depth tokens and adds a distillation loss that transfers the spatial perception of a dedicated depth model into the vision backbone, so no separate depth model has to be called at inference. Training-efficiency optimizations include Fully Sharded Data Parallel (FSDP), shard grouping for the action expert (reducing communication overhead), a mixed-precision strategy of bfloat16 storage plus float32 reduction, FlexAttention sparse multimodal attention, and torch.compile operator fusion — finally achieving near-linear throughput scaling from 8 to 256 GPUs.

03 实验Experiments

实验分为三部分：(1) GM-100 大规模真实世界基准——在 3 种机器人平台上各评测 100 项任务，指标为 Success Rate (SR) 和 Progress Score (PS)；(2) RoboTwin 2.0 仿真基准——干净场景与随机化场景；(3) 训练吞吐量分析。基线方法包括 WALL-OSS、GR00T N1.6 和 π₀.₅。The experiments have three parts: (1) the GM-100 large-scale real-world benchmark, evaluating 100 tasks on each of 3 robot platforms, with Success Rate (SR) and Progress Score (PS) as metrics; (2) the RoboTwin 2.0 simulation benchmark, with clean and randomized scenes; (3) training-throughput analysis. The baselines include WALL-OSS, GR00T N1.6 and π₀.₅.

GM-100 真实世界基准（Table 1）GM-100 real-world benchmark (Table 1)

平台Platform	方法Method	SR (%)	PS (%)
Agibot G1	WALL-OSS	2.99	8.75
Agibot G1	GR00T N1.6	5.23	12.63
Agibot G1	π₀.₅	7.77	21.98
Agibot G1	LingBot w/o depth	12.82	30.04
Agibot G1	LingBot w/ depth	11.98	30.47
AgileX	WALL-OSS	2.26	8.16
AgileX	GR00T N1.6	3.26	10.52
AgileX	π₀.₅	17.20	34.82
AgileX	LingBot w/o depth	15.50	36.31
AgileX	LingBot w/ depth	18.93	40.36
Galaxea R1Pro	WALL-OSS	6.89	14.13
Galaxea R1Pro	GR00T N1.6	14.29	24.83
Galaxea R1Pro	π₀.₅	14.10	26.14
Galaxea R1Pro	LingBot w/o depth	18.89	34.71
Galaxea R1Pro	LingBot w/ depth	20.98	35.40
平均Average	LingBot w/ depth	17.30	35.41

LingBot-VLA 平均 SR 超越 π₀.₅ 4.28%，平均 PS 超越 7.76%。LingBot-VLA exceeds π₀.₅ by 4.28% in average SR and by 7.76% in average PS.

RoboTwin 2.0 仿真基准（Table 2）RoboTwin 2.0 simulation benchmark (Table 2)

场景Setting	π₀.₅ SR (%)	LingBot w/o depth SR (%)	LingBot w/ depth SR (%)
干净场景 (Clean)Clean	82.74	86.50	88.56
随机化场景 (Randomized)Randomized	76.76	85.34	86.68

Scaling Law 与数据效率Scaling law and data efficiency

Pre-training data scaling law — 图5：预训练数据规模扩展实验。从 3,000 小时到 20,000 小时，Success Rate 和 Progress Rate 均呈现持续上升趋势，且"no signs of saturation"——说明更多数据仍有提升空间，scaling law 在多个平台上均成立。Figure 5: Pre-training data scaling experiment. From 3,000 hours to 20,000 hours, both Success Rate and Progress Rate keep rising and show "no signs of saturation" — indicating that more data still brings gains, and that the scaling law holds across multiple platforms.

Post-training data efficiency — 图6：后训练数据效率分析。仅用 80 个 demonstrations（相比 π₀.₅ 的完整 130 个）时，LingBot-VLA 在 SR 和 PS 上均超越使用全部数据的 π₀.₅，且两者性能差距随后训练数据量增加而显著拉大。Figure 6: Post-training data-efficiency analysis. With only 80 demonstrations (against the full 130 used by π₀.₅), LingBot-VLA already surpasses π₀.₅ trained on all the data in both SR and PS, and the performance gap widens markedly as the amount of post-training data grows.

训练吞吐量（Ablation）Training throughput (ablation)

基于 Qwen2.5-VL-3B-π 骨干，优化后的训练框架在 8-GPU 配置下达到 261 samples/second，较现有 VLA 专用代码库实现 1.5∼2.8× 提速。GPU 从 8 到 256 扩展时吞吐量近线性增长，验证了框架的大规模训练可行性。关键优化手段包括 FSDP 分布式训练、FlexAttention 稀疏注意力、torch.compile 算子融合及 bfloat16 混合精度。With the Qwen2.5-VL-3B-π backbone, the optimized training framework reaches 261 samples/second on an 8-GPU configuration, a 1.5∼2.8× speedup over existing VLA-specific codebases. Throughput grows near-linearly when scaling from 8 to 256 GPUs, confirming that the framework is viable for large-scale training. Key optimizations include FSDP distributed training, FlexAttention sparse attention, torch.compile operator fusion and bfloat16 mixed precision.

04 局限性Limitations

Note：论文未设专门的 Limitations 节；以下第 1 条为作者在正文中明确指出的问题，第 2、3 条为基于设计推断（inferred from the design）。Note: The paper has no dedicated Limitations section; item 1 below is a problem the authors state explicitly in the text, while items 2 and 3 are inferred from the design.

仅覆盖双臂机器人，平台多样性有限Covers dual-arm robots only; limited platform diversity

作者明确表示，未来工作将聚焦于"scaling the model versatility by integrating single-arm and mobile robotic data"——当前版本仅在 9 种双臂配置上训练，单臂与移动机器人场景尚未覆盖。（stated）The authors state explicitly that future work will focus on "scaling the model versatility by integrating single-arm and mobile robotic data" — the current version is trained on 9 dual-arm configurations only, and single-arm and mobile-robot scenarios are not yet covered. (stated)

仿真与真实环境的 sim-to-real gapThe sim-to-real gap between simulation and the real world

论文指出"simulation environments typically employ idealized physical models"，真实部署中摩擦、光照、传感器噪声等因素可能导致性能下降。（stated）The paper points out that "simulation environments typically employ idealized physical models"; in real deployment, friction, lighting, sensor noise and similar factors may degrade performance. (stated)

绝对成功率仍较低，泛化到长视野任务存在挑战Absolute success rates are still low, and generalizing to long-horizon tasks remains challenging

GM-100 最高平均 SR 仅为 17.30%，说明在百任务规模的开放式评测下模型仍有较大提升空间；测试集约 50% 的原子动作未出现在训练集 top-100 中，泛化难度较高。（inferred）The highest average SR on GM-100 is only 17.30%, showing that the model still has plenty of headroom under open-ended evaluation at hundred-task scale; about 50% of the atomic actions in the test set never appear in the training set top-100, which makes generalization harder. (inferred)