Obstacle-Avoidant Leader Following with a Quadruped Robot

Carmen Scheidemann, Lennart Werner, Victor Reijgwart, Andrei Cramariuc, Joris Chomarat, Jia-Ruei Chiu, Roland Siegwart, Marco Hutter

TL;DR

This paper introduces a robust "virtual leash" system that enables a quadruped robot (ANYmal) to autonomously follow a human leader through complex, crowded, and narrow environments. By fusing a custom radio-frequency (RF) Angle of Arrival sensor with vision and LiDAR, and pairing it with an optimized local obstacle avoidance planner, the system achieves resilient tracking and navigation without placing a heavy cognitive burden on the operator.

What Problem Does This Paper Address?

While autonomous human-following capabilities are common in commercial drones, ground-based robots face a much tougher challenge. Navigating dynamic environments like crowded hallways, narrow doorways, or busy industrial sites requires fast, reactive local path planning.

Existing ground-robot solutions typically rely on physical tethers, basic distance-keeping sensors, or purely visual tracking. These approaches break down in real-world scenarios: visual tracking fails during occlusions or in dense crowds, wireless beacons suffer from signal interference (multipathing), and most systems lack the smart obstacle avoidance needed to safely navigate around moving bystanders and static geometry. The authors tackle this by creating a highly redundant, multi-modal tracking and navigation pipeline that keeps the robot safely on the leader's heels.

Methods

The proposed architecture is divided into three core components: sensor fusion, leader tracking, and local navigation.

The complete pipeline architecture, spanning sensor fusion, EKF-based leader tracking, and Waverider-driven local navigation.
  1. Multi-Modal Sensor Fusion: The robot utilizes a Velodyne LiDAR, four RGB-D cameras, and a custom-built Angle of Arrival (AoA) sensor array. The human leader carries a pocket-sized RF transmitter. To identify the correct person in a crowd, the system runs YOLOv8 on the camera feeds to detect humans and cross-references their bounding boxes with the directional data from the AoA sensor.
  2. Leader Tracking: Once identified, the leader's position is tracked using an Extended Kalman Filter (EKF) with a linear motion model. The system projects the leader's point cloud forward in time, filters out environmental structures (like floors and walls) using RANSAC, and applies a KD-tree nearest-neighbor search to continuously associate new LiDAR points with the leader.
  3. Local Navigation (Waverider + RMPs): For obstacle avoidance, the authors adapted Waverider, a reactive path planner originally designed for UAVs. It integrates 3D occupancy mapping (wavemap) with Riemannian Motion Policies (RMPs). The navigation problem is solved at 50 Hz by combining multiple mathematical forces: an attractor policy pulling the robot toward the leader, a yaw policy keeping it facing the goal, and separate repulsor policies for static and dynamic obstacles.
The custom-built 360° Angle of Arrival (AoA) sensor array mounted on top of the robot.

Datasets

The paper does not introduce or utilize standard public datasets for benchmarking the entire system. Instead, the authors validated their pipeline through extensive real-world deployments and custom experimental setups using the ANYmal quadruped platform in both indoor and outdoor environments.

Key Insights and Ideas

  • Decoupling Static and Dynamic Obstacles: Standard 3D occupancy mapping frameworks assume the world is static; trying to adapt them for dynamic objects usually introduces latency or map noise. The authors elegantly sidestep this trade-off by using visual detections (YOLOv8) to explicitly filter out humans from the static map, handling them instead via a dedicated dynamic obstacle RMP policy.
  • RF Beacon as an Identity Anchor: Purely visual tracking easily loses the target during brief occlusions or when another person walks between the robot and the leader. The custom RF AoA sensor acts as an un-occludable identity anchor, allowing the EKF to maintain a reliable heading estimate and instantly re-acquire the leader visually once the line of sight returns.

Key Results

  • AoA Accuracy: In open spaces, the custom AoA sensor achieved a mean error of just 3° (with an 11° standard deviation). In a highly challenging multipath environment (standing directly next to a wall), the mean error rose to a still-manageable 7°.
  • Computational Efficiency: The entire local navigation and replanning pipeline runs at a constant 50 Hz while consuming only 2.5 CPU cores on the robot's onboard computer.
  • Robust Behavior: In real-world testing, the robot successfully navigated around static obstacles like chairs and benches while tracking the leader. When dynamic actors intentionally blocked the robot's path or caused temporary tracking swaps, the system consistently self-corrected within seconds using the AoA beacon.
Visualization of the robot's path (arrow clusters) and the attractive/repelling forces (green/red) computed by the Waverider policy over time as it follows the leader (purple line).

Limitations and Future Work

While highly robust, the custom AoA sensor's performance degrades slightly (mean error increases to 4° and standard deviation to 14°) if the beacon moves significantly out of the horizontal plane relative to the antenna array.

Additionally, the current obstacle avoidance policies operate within $SE(2)$ (planar translation and yaw rotation). Because the robot does not actively plan changes to its roll, pitch, or height via the RMP framework, it cannot yet autonomously decide to step over high obstacles or duck under low overhangs. The authors note that extending the policies to $SE(3)$ is a promising direction for future work.

Summary updated

Abstract

Personal mobile robotic assistants are expected to find wide applications in industry and healthcare. For example, people with limited mobility can benefit from robots helping with daily tasks, or construction workers can have robots perform precision monitoring tasks on-site. However, manually steering a robot while in motion requires significant concentration from the operator, especially in tight or crowded spaces. This reduces walking speed, and the constant need for vigilance increases fatigue and, thus, the risk of accidents. This work presents a virtual leash with which a robot can naturally follow an operator. We use a sensor fusion based on a custom-built RF transponder, RGB cameras, and a LiDAR. In addition, we customize a local avoidance planner for legged platforms, which enables us to navigate dynamic and narrow environments. We successfully validate on the ANYmal platform the robustness and performance of our entire pipeline in real-world experiments.
BibTeX
@inproceedings{scheidemann2025obstacle,
  author = {Carmen Scheidemann and Lennart Werner and Victor Reijgwart and Andrei Cramariuc and Joris Chomarat and Jia-Ruei Chiu and Roland Siegwart and Marco Hutter},
  title = {{Obstacle-Avoidant Leader Following with a Quadruped Robot}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2025},
  doi = {10.1109/icra55743.2025.11127912},
}
Figure