Clutter-Aware Spill-Free Liquid Transport via Learned Dynamics

TL;DR

This work tackles spill-free transport of open-top liquid containers in cluttered environments by marrying geometry-informed planning with learned spillage dynamics. It introduces Spill-Free RRT* (SFRRT*), which expands the action space by allowing high tilt angles and full 3D orientation, using an informed sampler derived from container geometry to guide sampling, and a transformer-based Spill-Free Classifier (SFC) to assess trajectories. A Spill-Free Time Parameterization (SFTP) then iteratively adjusts trajectory jerk under Ruckig constraints to ensure spill-free motion, guided by SFC predictions. Training on 700 real trajectories enables generalization across container shapes and fill levels, with real-world experiments showing substantially larger feasible orientation ranges and robust obstacle avoidance. The approach advances practical robotic liquid handling by avoiding continuous liquid-surface monitoring and enabling adaptable, spill-free operations in complex environments.

Abstract

In this work, we present a novel algorithm to perform spill-free handling of open-top liquid-filled containers that operates in cluttered environments. By allowing liquid-filled containers to be tilted at higher angles and enabling motion along all axes of end-effector orientation, our work extends the reachable space and enhances maneuverability around obstacles, broadening the range of feasible scenarios. Our key contributions include: i) generating spill-free paths through the use of RRT* with an informed sampler that leverages container properties to avoid spill-inducing states (such as an upside-down container), ii) parameterizing the resulting path to generate spill-free trajectories through the implementation of a time parameterization algorithm, coupled with a transformer-based machine-learning model capable of classifying trajectories as spill-free or not. We validate our approach in real-world, obstacle-rich task settings using containers of various shapes and fill levels and demonstrate an extended solution space that is at least 3x larger than an existing approach.

Figures (10)

• Figure 1: Motivated by the need for robots to perform in cluttered environments such as a chemistry setup where robots must transport chemistry vessels inside a fume hood (top row), this paper targets the challenge of enabling spill-free liquid transport while avoiding obstacles. Our approach introduces two key enhancements over existing methods: (i) increasing the allowable tilt angle to the maximum quasi-static spill-free tilt angle, rather than the limited angles used previously pendulumpendulum2019 (second row); (ii) enabling rotation about all axes throughout the trajectory, instead of maintaining fixed yaw or orientation (third row) pendulumspringdamper2021.
• Figure 2: System diagram. The inputs are the start, goal, and obstacles, and the output is the trajectory. First, (i) a spill- and collision-free path is generated using RRT* and the informed sampler. The informed sampler utilizes the container properties to avoid sampling states such as an upside-down container. Then, (ii) a spill- and collision-free trajectory is generated using a time parameterization algorithm alongside the spill-free classifier model. The SFC model classifies trajectories as either spill-free or not and is utilized to set the jerk of the trajectory, ensuring the generation of a spill-free trajectory.
• Figure 3: (a) depicts modeling containers as frustum of cones. Two cases occur when tilting the container as illustrated in (b) and (c). Case I is when the liquid forms both a trapezoid and a triangle shape. Case II shows when the liquid forms only a triangular shape. In both cases, $\theta_t$ represents the maximum tilt angle being calculated.
• Figure 4: Spill-Free Classifier Model Architecture The input is the trajectory and the container properties. The output is a boolean stating whether the trajectory is spill-free or not.
• Figure 5: The two scenarios from \ref{['table:tasks']}: The original colored robot for start position, green for goal. Blue indicates intermediate robot states.