Pose Estimation of a Thruster-Driven Bioinspired Multi-Link Robot

TL;DR

This work addresses pose estimation for a free-floating, underactuated, multi-link robot driven by thrusters and equipped with a single gyroscope per link. It develops a quasi-static, drag-dominated dynamic model and a gyroscope-based measurement model, then employs an Enhanced-GP UKF to learn residuals and fuse IMU data for offline pose estimation using two gait training regimes. The offline experiments with LandSalp show that a GP-UKF trained on a multi-gait dataset generalizes well to forward, backward, left, and right gaits, with comparable accuracy to a forward-gait-specific filter, despite non-Gaussian noise and drift inherent to dead-reckoning. The results highlight the potential for reduced data requirements and improved generalizability across gaits, while also outlining the path toward real-time onboard deployment and scaling to longer, more complex multi-link systems.

Abstract

This work demonstrates pose (position and shape) estimation for a free-floating, bioinspired multi-link robot with unactuated joints, link-mounted thrusters for control, and a single gyroscope per link, resulting in an underactuated, minimally sensed platform. Through a proof-of-concept hardware experiment and offline Kalman filter analysis, we show that the robot's pose can be reliably estimated. State estimation is performed using an unscented Kalman filter augmented with Gaussian process residual learning to compensate for non-zero-mean, non-Gaussian noise. We further show that a filter trained on a multi-gait dataset (forward, backward, left, right, and turning) performs comparably to one trained on a larger forward-gait-only dataset when both are evaluated on the same forward-gait test trajectory. These results reveal overlap in the gait input space, which can be exploited to reduce training data requirements while enhancing the filter's generalizability across multiple gaits.

Figures (8)

• Figure 1: Schematic comparison of the LandSalp robot and the corresponding sea salp with linear-like architectures. There are two passive joints connecting the links. The orientations ($\beta_i$) of the three motor-omniwheel assemblies used to generate jet thrust and viscous drag are $-57^\circ$, $-130^\circ$, and $-57^\circ$, respectively, from left to right, relative to the link axis.
• Figure 2: Control trajectories (left) and planned shape trajectories (right) for the forward, backward, left, right, and turning gaits.
• Figure 3: Residual distributions of the process model (blue) and measurement model (orange), obtained by comparing the model predictions with motion capture and gyroscope measurements. The dashed black line denotes zero, and the dashed red line denotes the sample mean. Residuals were computed from the forward gait training set at 200 Hz.
• Figure 4: Snapshots of the forward motion of the LandSalp experiment. The first panel presents the experimental trajectories together with robot base frames. The remaining five panels illustrate the robot’s positions at approximately the quarter intervals of the entire motion cycle, as well as the configurations corresponding to the quarter intervals of the gait cycle. White arrows indicate the desired direction of velocity.
• Figure 5: LandSalp experiment setup. The robot runs on a platform integrated with the OptiTrack motion capture system. Motion capture data is processed by a desktop computer and transmitted via the Robot Operating System (ROS) to a laptop that executes the control algorithm. The robot is connected by two cables, one for power supply and the other for command signals.