NeuRIO: Physics-Guided Edge-State Graph Learning for Zero-Shot Multi-Robot Relative Inertial Odometry NeuRIO：面向多机器人零样本相对惯性里程计的物理引导边状态图学习

Zhehan Li, …
Zhejiang University · Under review at IEEE ICRA 浙江大学 · IEEE ICRA 在投
First author 第一作者
Full author list will be released with the preprint. 完整作者列表将随预印本一同公开。
Loading the replay… 正在加载回放…
drag to orbit · scroll to zoom · double-click to reset 拖动旋转 · 滚轮缩放 · 双击复位
0.00 s

What you are looking at. Frame-by-frame replay of the frozen, simulation-trained model on real robot sequences. The observer is one robot; for every other robot the hollow circle is the motion-capture ground truth, the filled dot is the estimate, and the line between them is that frame’s position error. Estimates are produced in the observer’s gravity-aligned frame and drawn unmodified — the world view additionally uses the observer’s ground-truth pose to place the formation back in the room, so it is for reading the geometry, not for judging accuracy. The line to each robot shows what arrived that frame: solid = bearing + range, dashed = bearing only, dotted = range only, and no line at all means the pose is carried by inertial propagation and the model’s own memory. The dashed circle is the uncertainty σ the model reports for itself; it never enters the error numbers. 这里展示的是什么。冻结参数、仅用仿真训练的模型在真机序列上的逐帧回放。观测者是其中一台机器人；对其余每台机器人，空心圆是动捕真值，实心点是估计，两点之间的连线长度就是该帧的位置误差。估计输出在观测者的重力对齐坐标系下，图中原样绘制；「世界系俯视」额外用观测者的真值位姿把编队拼回房间，因此它用来看几何关系，而不用来判断精度。指向每台机器人的连线表示这一帧收到了什么观测：实线 = 方位 + 测距，虚线 = 只有方位，点线 = 只有测距，没有连线则说明该帧完全靠惯性递推与模型自身的记忆维持。虚线圆是模型自报的不确定度 σ，不参与任何误差计算。

One simulation-trained estimator, deployed zero-shot across real robot teams. 一个仅用仿真训练的估计器，零样本部署到不同真机集群。

Overview 项目概述

Multi-robot relative localization must combine asynchronous bearing, UWB range and inertial streams despite changing visibility and persistent sensor failures. NeuRIO represents each robot with a gravity-aligned node state and each directed robot pair with its own persistent edge state. These edge states condition attention and carry pair-specific visibility and corruption history across keyframes, allowing a single causal network to stream the full 6-DoF relative pose of every neighbor. The model is trained on 400 simulated trajectories with coupled camera, ranging and inertial failure modes; its parameters are then frozen and transferred directly to real robots. Across 24 real sequences, the zero-shot model achieves a mean position error of 0.121 m and a mean rotation error of 2.53°. An optional label-free stage adapts a low-rank input module using geometric measurement residuals rather than pose labels.

多机器人相对定位需要在可见关系不断变化、传感器故障具有持续性的条件下，融合异步的方位、UWB 测距与惯性数据。NeuRIO 为每个机器人维护重力对齐的节点状态，并为每个有向机器人对维护独立、可跨关键帧传递的边状态；边状态既调制注意力，也记录该机器人对的可见性与观测异常历史，使一个因果图网络能够流式输出全部邻居的 6-DoF 相对位姿。模型在包含相机、测距与惯性耦合故障的 400 条仿真轨迹上训练，随后冻结参数并直接部署到真机。零样本模型在 24 段真机序列上的平均位置误差为 0.121 m、平均旋转误差为 2.53°。此外，可选的无标签阶段利用几何测量残差适配低秩输入模块，无需位姿真值。

Key Contributions 主要贡献

  • Persistent node–edge state. Separate causal states for robots and ordered robot pairs preserve pair identity and measurement history as visibility changes.
  • Physics-guided sim-to-real transfer. Gravity-aligned coordinates and a failure-mode simulator encode known structure while exposing the model to occlusion, non-line-of-sight ranging and persistent device faults.
  • Optional label-free adaptation. A low-rank input adapter is updated from bearing, range and inertial-preintegration residuals, while the source network remains frozen.
  • Real-world accuracy and throughput. The frozen model averages 0.121 m / 2.53° over 24 real sequences; on the reported RTX 4070 Ti SUPER benchmark, median latency is 1.9 ms at 10 robots and the 99th percentile remains below 20 ms up to 200 robots.
  • 持续更新的节点—边状态。分别为机器人和有向机器人对维护因果状态，在可见关系变化时保留机器人对身份与观测历史。
  • 物理引导的 Sim-to-Real。重力对齐坐标系注入几何结构，故障模式仿真器则覆盖遮挡、非视距测距和持续性器件故障。
  • 可选的无标签适配。保持源网络冻结，仅依据方位、测距与惯性预积分残差更新低秩输入适配器。
  • 真机精度与吞吐。冻结模型在 24 段真机序列上平均达到 0.121 m / 2.53°；在论文所用 RTX 4070 Ti SUPER 上，10 台机器人时中位延迟为 1.9 ms，扩展至 200 台时第 99 百分位延迟仍低于 20 ms。