
ICRA 2026|KAN We Flow?(机器人控制)
Quick Take
The KAN-We-Flow model enhances robotic manipulation through 3D flow matching, utilizing RWKV and KAN to reduce parameters by 86.8% while achieving superior performance on Adroit, Meta-World, and DexArt benchmarks. This approach supports real-time control with inference times of 8-11ms, addressing long-sequence dependencies efficiently.
Key Points
- KAN-We-Flow achieves state-of-the-art performance on multiple robotic benchmarks.
- Parameter count reduced to approximately 33.6M, significantly lowering computational costs.
- Inference time of 8-11ms enables 100Hz real-time control.
- RWKV and GroupKAN enhance model efficiency and performance stability.
- ACR regularization notably decreases action drift in long prediction windows.
Article Excerpt
From source RSS / original summary原文作者:用户“多多的贾维斯”原文链接:https://www. xiaohongshu. com/ KAN We Flow? Advancing Robotic Manipulation with 3D Flow Matching via KAN & RWKV 一、整体概述1. 本文提出KAN-We-Flow,一种用于机器人三维操作的高效流匹配策略模型。 2. 核心贡献在于用RWKV与KAN替代传统大规模UNet骨干,在保持甚至提升成功率的同时,大幅降低参数量与推理延迟。 3. 方法在Adroit、Meta-World、DexArt三大基准上取得当前最优或并列最优性能,参数量减少约86. 8%,支持实时控制。
二、研究背景1、扩散式策略① 优点是动作分布建模能力强② 缺点是需要多步去噪,推理慢、模型重,不利于真实机器人部署2、流匹配策略① 通过学习一步向量场实现快速生成② 但现有方法仍大量依赖UNet,计算与存储开销依旧偏大3、核心问题如何在保证精度的前提下,进一步压缩模型规模并提升实时性 三、动机直觉1、RWKV具备线性复杂度的时序建模能力,适合长时序动作预测2、KAN基于可学习的一维函数逼近,能以更少参数表达复杂非线性映射3、将二者结合,有望同时解决“长时序依赖”和“参数效率”问题四、技术路线1、整体框架① 采用一致性流匹配,实现一步动作生成② 输入为点云感知、机器人状态与时间编码2、核心网络① RWKV-KAN骨干网络* RWKV负责时间与通道混合,建模动作序列上下文* GroupKAN对特征通道进行分组的非线性函数校准,替代传统MLP② Action Consistency Regularization(ACR)* 通过欧拉外推,将一步预测动作与专家轨迹在末端对齐* 提供额外监督,稳定训练,不增加推理成本3、学习目标联合一致性流匹配损失与ACR正则项进行端到端训练五、实验结果1、性能表现① 在Adroit、Meta-World、DexArt上整体成功率优于FlowPolicy与DP3② 在高难度、长时序任务中优势更明显2、效率对比① 参数量约33.
6M,相比DP3减少86. 8%② 推理时间约8–11ms,满足100Hz实时控制3、消融实验① RWKV、GroupKAN与ACR均对性能有稳定增益② ACR在长预测窗口下显著降低动作漂移 雷峰网
Reader Mode unavailable (could not extract clean content).
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from 雷峰网机器人
See more →
ICRA 2026 | 港中文高源、林天麟团队提出自发共适应策略:基于元学习赋能异构多机器人系统协同进化
The Hong Kong Polytechnic University team presented a meta-learning framework for heterogeneous multi-robot systems at ICRA 2026, enhancing adaptability in dynamic environments. Their approach improved task completion efficiency by 21% and reduced human burden by 39%, demonstrating effective human-robot interaction.


