Uncertainty-aware Planning with Inaccurate Models for Robotized Liquid Handling
Marco Faroni, Carlo Odesco, Andrea M. Zanchettin, Paolo Rocco · 2025
Read Uncertainty-aware Planning with Inaccurate Models for Robotized Liquid Handling by Marco Faroni, Carlo Odesco, Andrea M. Zanchettin, Paolo Rocco (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-20861.txt.
Introduction
Marco Faroni and co-authors present, in a 2025 paper, a planning method for a robot that pours liquid, when the model guiding it is known to be imperfect. Physics simulators and learned models are vital for hard robot tasks such as handling soft objects and liquids. They often miss, due to gaps in knowledge or the shift from simulation to the real world. Pouring is a sharp case, since a model trained on few examples can fail in new setups.
Their planner is a Monte Carlo tree search that knows about its own doubt. It estimates the model's uncertainty for each candidate action, and steers the search toward actions with lower predicted doubt. On a liquid pouring task, success rates improved even with models trained on minimal data. The method beat traditional planners, and the authors offer it for robust robot decisions more widely.
Catalog record
| CREATOR | Marco Faroni, Carlo Odesco, Andrea M. Zanchettin, Paolo Rocco |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-20861
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Uncertainty-aware Planning with Inaccurate Models for Robotized Liquid Handling Marco Faroni, Carlo Odesco, Andrea M. Zanchettin, Paolo Rocco Abstract Physics-based simulations and learning-based models are vital for complex robotics tasks like deformable object manipulation and liquid handling. However, these models often struggle with accuracy due to epistemic uncertainty or the sim-to-real gap. For instance, accurately pouring liquid from one container to another poses challenges, particularly when models are trained on limited demonstrations and may perform poorly in novel situations. This paper proposes an uncertainty-aware Monte Carlo Tree Search (MCTS) algorithm designed to mitigate these inaccuracies. By incorporating estimates of model uncertainty, the proposed MCTS strategy biases the search towards actions with lower predicted uncertainty. This approach enhances the reliability of planning under uncertain conditions. Applied to a liquid pouring task, our method demonstrates improved success rates even with models trained on minimal data, outperforming traditional methods and showcasing its potential for robust decision-making in robotics. I Introduction Physics-based simulations and learning-based models are extensively used in robotics to perform complex tasks such as deformable object manipulation [ 1 , 2 , 3 , 4 , 5 ] , contact-rich manipulation [ 6 , 7 , 8 ] , control of soft robots [ 9 , 10 ] , and liquid handling [ 11 , 12 ] . These models are often inaccurate in predicting the outcome of actions (e.g., because of the epistemic uncertainty of learned models or the sim-to-real gap of physics simulators). A notable example is the liquid handling scenario in Fig. 1 : A robot has to pour liquid from one container to another until it reaches a certain level. Ideally, one could accurately model the quantity of poured liquid based on the robot’s tool motion (e.g., via fluid simulation), but this requires a deep understanding of the geometry and dynamics properties of the problem or the collection of large experimental datasets. Suppose a pouring model is built from a handful of pouring demonstrations. The model will be very inaccurate, especially for states and actions that are far from those seen in the training set. Fig. 1: A robotized pouring task: the robot arm shall pour liquid from one container to another until it reaches a reference level. If an estimate of the inaccuracy of the model predictions is available, this information can be embedded in the planning algorithm to avoid inaccurate states and actions [ 3 , 13 ] . Interestingly, machine learning techniques commonly used for dynamics modeling in robotics come with an estimate of the model epistemic uncertainty at a given point. For example, the variance of a Gaussian Process (GPs) is often seen as a proxy of the prediction uncertainty. Similarly, the variance of the predictions of Neural Network ensembles measures the consensus of the networks: if the variance is large, the average prediction is likely to be inaccurate. This work proposes an uncertainty-aware Monte Carlo Tree Search (MCTS) algorithm that uses such uncertainty information to cope with model inaccuracy. We use the uncertainty value of each prediction to inform the MCTS search and bias it towards transitions with small uncertainty. As a result, the planner is likely to choose accurate states and actions, leading to more accurate task executions. We demonstrate the approach with a liquid pouring problem. To do so, we model pourings with Gaussian Process Regression (GPR) and use the variance of GPs as an accuracy index for the model predictions. Then, we use the uncertainty-aware MCTS to choose the sequence of pouring actions to reach a given filling level. We show that our approach performs pouring tasks even when the model uncertainty is very large (e.g., a GP with a handful of points), leading to a much higher success rate than the baselines. II Related Works Planning under uncertainty gathers all those planning methods that explicitly reason about the difference between the model and the true system. The goal of these methods is to find a sequence of actions that brings the system from an initial to a final state even though the model and the true dynamics differ. If such a difference is due to the inherent randomness of the system (i.e., aleatoric uncertainty), a common approach is to propagate the uncertainty on the search tree. The planner will search for a robust solution that reaches the goal despite the transitions’ randomness [ 14 , 15 , 16 ] . If the model error is caused by the limited dataset or the lack of expressivity of the model, we are dealing with epistemic uncertainty. This last case is less studied and is the focus of this work. In the realm of sampling-based motion planning, existing approaches estimate a Model Deviation Estimate ( MDE ) predicting the accuracy of a transition and use it to bias the search. Mitrano et al. [ 3 ] train a classifier to deem state-and-action pairs accurate or inaccurate during the search. Inaccurate transitions are not added to the tree in a sampling-based path planner. Faroni et al. [ 17 ] use the MDE as a cost function within an asymptotically optimal planner so that the resulting path should minimize the cumulative model mismatch. Other approaches limit the search space to a trusted region after an offline analysis of the model performance [ 18 ] . All these approaches rely on a computationally expensive training phase to learn which actions can be trusted and use such information in a general motion planning algorithm. Considering graph-based search, Vemula et al. [ 19 , 20 ] use A ∗ to plan robot movements and, after the execution of each action, assign a high cost to transitions whose deviation from the model is larger than a given threshold. MCTS is becoming popular in robotics planning. Recent applications include path planning with Partially Observable Markov Decision Processes (POMDP) [ 21 , 22 ] , multi-robot task allocation [ 23 ] and path planning [ 24 ] , and rearrangement planning [ 25 ] . Existing works typically leverage MCTS to account for multi-agent decision-making, by modeling the multi-agency as a POMDP and using MCTS to solve the resulting problem. Unlike previous works, ours uses MCTS to explicitly reason through model uncertainty in the context of robotics planning. Compared to existing sampling-based and graph-based planners, MCTS can be easily extended to account for uncertainty [ 26 ] . Specific to robotized liquid pouring, most existing works focus on accurate modeling of the liquids’ dynamics via deep learning or fluid simulation. For example, Pan et al. [ 27 ] combined a fine-grained fluid simulation with an optimization-based planner to achieve the pouring task. The motion planning problem was formulated as a continuous numerical optimization problem in the high-dimensional robot trajectory space, following an objective function that accounts for object avoidance, smoothness, and a target goal that ensures fluid particles enter the target container. Similarly, Schober et al. [ 28 ] segment the image of the container, simulate possible pourings and select the best action to perform in the real-world. Zhang et al. [ 29 ] used imitation learning to learn pouring tasks from human demonstrations. They devised a one-shot domain adaptation technique to reduce the sim-to-real gap between simulations and