CronusVLA

01 动机Motivation

现有基于预训练视觉-语言模型（VLM）的 VLA 模型仍受限于单帧图像范式，无法充分利用多帧历史提供的时序信息。 直接将多帧图像输入 VLM backbone 会因 self-attention 的二次方计算复杂度带来巨大的计算开销和推理延迟。 低级策略（low-level policies）的研究已表明多帧历史观测可提升性能与鲁棒性，但如何将其高效地引入大型 VLA 模型仍是未解难题。Existing VLA models built on pretrained vision-language models (VLM) remain constrained by the single-frame image paradigm and cannot fully exploit the temporal information offered by multi-frame histories. Feeding multiple frames directly into a VLM backbone incurs substantial computational overhead and inference latency because of the quadratic complexity of self-attention. Work on low-level policies has shown that multi-frame historical observations improve performance and robustness, yet efficiently introducing them into large VLA models remains an open problem.

"these models remain constrained by the single-frame image paradigm and fail to fully leverage the temporal information offered by multi-frame histories, as directly feeding multiple frames into VLM backbones incurs substantial computational overhead and inference latency."

02 方法Method

CronusVLA 采用两阶段训练策略：第一阶段在大规模具身数据集上进行单帧 VLA 预训练，建立有效的具身视觉-语言基础； 第二阶段通过 feature chunking 将 VLM backbone 的输出从离散 token 替换为可学习特征（learnable features）， 并经由跨帧解码器（cross-frame decoder）汇聚历史多帧信息，同时利用队列机制（queue mechanism）实现快速推理。CronusVLA adopts a two-stage training strategy: the first stage performs single-frame VLA pretraining on large-scale embodied datasets, establishing an effective embodied vision-language foundation; the second stage uses feature chunking to replace the discrete-token output of the VLM backbone with learnable features, aggregates multi-frame history through a cross-frame decoder, and exploits a queue mechanism for fast inference.

单帧预训练（Single-Frame Pretraining）Single-Frame Pretraining

第一阶段在大规模异构具身数据集上训练基础单帧 VLA 模型，采用自回归方式预测离散动作 token（量化为 256 bins）。 该阶段保留了 VLM backbone 的视觉感知能力，为后续多帧扩展奠定强基础。 模型配置包括：CronusVLA-7B（Llama 2 7B backbone，DINOv2 + SigLIP 视觉编码器） 与 CronusVLA-0.5B（Qwen2.5 0.5B backbone）。The first stage trains a base single-frame VLA model on large-scale heterogeneous embodied datasets, predicting discrete action tokens (quantized into 256 bins) autoregressively. This stage preserves the visual perception ability of the VLM backbone and lays a strong foundation for the subsequent multi-frame extension. Model configurations include CronusVLA-7B (Llama 2 7B backbone, DINOv2 + SigLIP visual encoders) and CronusVLA-0.5B (Qwen2.5 0.5B backbone).

多帧后训练与跨帧解码器（Multi-Frame Post-Training & Cross-Frame Decoder）Multi-Frame Post-Training & Cross-Frame Decoder

后训练阶段将 VLM backbone 的预测目标从离散 token 切换为连续的可学习特征，并通过 feature chunking 将 M 步历史帧的特征组织为结构化序列送入跨帧解码器。 跨帧解码器基于 DiT（Diffusion Transformer）设计，包含：In post-training the prediction target of the VLM backbone switches from discrete tokens to continuous learnable features, and feature chunking organizes the features of M historical steps into a structured sequence fed to the cross-frame decoder. The cross-frame decoder is built on DiT (Diffusion Transformer) and contains:

• Cross-attention 机制：计算复杂度与帧数线性增长（区别于直接输入多帧图像的二次方复杂度）Cross-attention mechanism: computational cost grows linearly with the number of frames (in contrast to the quadratic complexity of feeding multiple frames directly)
• Feature modulator：动态平衡当前帧与历史帧特征的贡献Feature modulator: dynamically balances the contributions of the current frame and the historical frames
• 扩散损失（Diffusion loss）：用于动作序列预测Diffusion loss: used for action sequence prediction

历史帧特征采用 stop-gradient 操作，防止梯度更新 VLM backbone，从而在保留单帧感知能力的同时实现解码器层面的时序建模（multi-frame regularization）。 推理时利用 feature 缓存（feature caching）消除重复的视觉-语言计算，实现 8.73 Hz 的高速推理（基线仅 3.09 Hz，提升 184%）。 后训练数据集为 Bridge-v2 + Fractal（148k 轨迹，5M 多帧片段）。Historical-frame features use a stop-gradient operation that prevents gradients from updating the VLM backbone, so temporal modeling is realized at the decoder level while single-frame perception is preserved (multi-frame regularization). At inference, feature caching eliminates redundant vision-language computation and reaches 8.73 Hz (the baseline reaches only 3.09 Hz, a 184% improvement). The post-training dataset is Bridge-v2 + Fractal (148k trajectories, 5M multi-frame clips).

SimplerEnv-OR 鲁棒性基准SimplerEnv-OR Robustness Benchmark

为评估时序与空间干扰下的鲁棒性，论文提出 SimplerEnv-OR benchmark，包含 24 种观测干扰类型（模糊、抖动、丢帧、遮挡、噪声等）与 120+ 严重程度等级，覆盖 Global、Local、Discrete 三大类别，支持空间维度与时序维度的独立评估。To evaluate robustness under temporal and spatial disturbances, the paper introduces the SimplerEnv-OR benchmark, which covers 24 types of observational disturbance (blur, jitter, frame drop, occlusion, noise, etc.) and more than 120 severity levels across the three categories Global, Local and Discrete, supporting independent evaluation along the spatial and the temporal dimension.

03 实验Experiments

实验在三种具身平台（Google Robot VM/VA、WidowX、Franka）上进行，涵盖 SimplerEnv 仿真 benchmark、LIBERO benchmark、SimplerEnv-OR 鲁棒性评测及真实世界 Franka 机器人实验。 基线包括 OpenVLA、TraceVLA、RoboVLMs、SpatialVLA、DP3 等。Experiments are run on three embodiments (Google Robot VM/VA, WidowX, Franka) and cover the SimplerEnv simulation benchmark, the LIBERO benchmark, the SimplerEnv-OR robustness evaluation and real-world Franka robot experiments. Baselines include OpenVLA, TraceVLA, RoboVLMs, SpatialVLA and DP3.

SimplerEnv 仿真性能（Table 1）SimplerEnv Simulation Performance (Table 1)

相较 TraceVLA，CronusVLA-7B 在 Google VM 上取得 +71.6% 的相对提升，Google VA 上提升 +48.2%。Compared with TraceVLA, CronusVLA-7B achieves a relative gain of +71.6% on Google VM and +48.2% on Google VA.

LIBERO 基准测试（Table 2）LIBERO Benchmark (Table 2)

长时序任务（LIBERO-Long）成功率从 OpenVLA 的 53.7% 提升至 94.0%（+40.3%）；整体平均成功率较 OpenVLA 提升 +26.8%（论文摘要原文数据）。On long-horizon tasks (LIBERO-Long) the success rate rises from OpenVLA's 53.7% to 94.0% (+40.3%); the overall average success rate improves by +26.8% over OpenVLA (figure as stated in the paper's abstract).

SimplerEnv-OR 鲁棒性评测（Table 3）SimplerEnv-OR Robustness Evaluation (Table 3)

真实世界 Franka 实验Real-World Franka Experiments

消融实验（Ablations）Ablations

消融实验关键结论：Key conclusions of the ablation study:

• 仅加入多帧建模（+M.F.）：成功率 +4.5%，速度 -40.3%Multi-frame modeling only (+M.F.): success rate +4.5%, speed -40.3%
• 加入 decoder 与连续特征（+M.F. +Dec.）：成功率 +48.2%，速度提升 2.8×Adding the decoder and continuous features (+M.F. +Dec.): success rate +48.2%, speed improved 2.8×
• 多帧正则化（multi-frame regularization，stop-gradient）：加速收敛并提升整体性能Multi-frame regularization (stop-gradient): faster convergence and better overall performance

04 局限性Limitations