Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials
Cenk Cetin, Shreyas Pouli, Gabriella Pizzuto · 2026
Read Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials by Cenk Cetin, Shreyas Pouli, Gabriella Pizzuto (2026) 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-2603-10979.txt.
Introduction
Cenk Cetin, Shreyas Pouli, and Gabriella Pizzuto teach a robot arm to scrape samples from vial walls, a fiddly lab task. Materials range from granular and powdery to viscous. Their framework pairs a low-level impedance controller, which keeps contact stable and compliant, with a reinforcement learning agent. The agent adjusts the end-effector force using perception feedback about the material's location. Trained in simulation with per-particle dislodgement thresholds, the policy transfers to a real robot. It beats a fixed-force baseline across five material setups.
Catalog record
| CREATOR | Cenk Cetin, Shreyas Pouli, Gabriella Pizzuto |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39283 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.10979
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning Adaptive Force Control for Contact-Rich Sample Scraping with Heterogeneous Materials Thanks: School of Computer Science & Informatics, University of Liverpool, UK. Thanks: Department of Chemistry, University of Liverpool, UK. Thanks: This work was supported by the Royal Academy of Engineering under the Research Fellowship Scheme and Google DeepMind Research Ready programme. Cenk Cetin, Shreyas Pouli, Gabriella Pizzuto Abstract The increasing demand for accelerated scientific discovery, driven by global challenges, highlights the need for advanced AI-driven robots. Robotic systems in this area are still in their infancy, often being highly specialised and hand-engineered. This limits their application in early-stage materials discovery, where sample properties are heterogeneous and unpredictable. Deploying robotic chemists in human-centric labs is key for the next horizon of autonomous discovery, as complex tasks still demand the dexterity of human scientists. Robotic manipulation in this context is uniquely challenged by handling diverse materials (from granular to cohesive), under varying lab conditions. For example, humans effortlessly use spatulas for scraping materials from vial walls. Automating this process is challenging because it goes beyond simple robotic insertion tasks and traditional lab automation, requiring the execution of fine-grained movements within a constrained environment (the sample vial). Our work proposes an adaptive control framework to address this, relying on a low-level Cartesian impedance controller for stable and compliant physical interaction and a high-level reinforcement learning agent that learns to dynamically adjust interaction forces at the end-effector. The agent is guided by perception feedback, which provides the material’s location. We first created a task-representative simulation environment with a Franka Research 3 robot, a scraping tool, and a sample vial containing heterogeneous materials. To facilitate the learning of an adaptive policy and model diverse characteristics, the material is modelled as a collection of spheres, where each sphere is assigned a unique dislodgement force threshold, which is procedurally generated using Perlin noise. We train an agent to autonomously learn and adapt the optimal contact wrench for a sample scraping task in simulation and then successfully transfer this policy to a real robotic setup. Our method was evaluated across five different material setups, outperforming a fixed-wrench baseline by an average of 10.9%. Fig. 1 : The proposed control architecture for autonomous sample scraping. A high-level RL policy receives a visual state ( s_{\text{p}}=[v_{1},v_{2},v_{3}]^{T} ), with v_{i}=[c_{ix},c_{iy},c_{iz},p_{i}]^{T} denoting each cluster’s centroid and residue percentage) and the robot’s Cartesian state with the external wrench ( x,\dot{x},F_{\text{ext}} ). The policy outputs a hybrid action command ( \boldsymbol{a}_{t}=[f_{x}^{c},\tau_{y}^{c},z^{D}]^{T} ) at 10 Hz, defining a desired force, torque, and position in the z-axis. This command is used as a goal to the Cartesian impedance controller running at 500 Hz, which generates the compliant joint torques ( \tau_{c} ). I Introduction Transforming materials discovery is crucial for tackling topical issues [ 25 ] , from clean energy to pharmaceuticals. Traditionally, this process has been slow and costly due to its heavy reliance on manual, repetitive, and often hazardous experimental workflows performed by human chemists. The subjective nature of these manual experiments has historically led to inconsistencies, contributing to a reproducibility crisis in the field [ 27 ] . To accelerate fundamental breakthroughs, there is a growing paradigm shift towards using intelligent autonomous systems, with a diverse array of general-purpose robots being deployed to automate workflows like synthesis [ 6 ] and photocatalysis [ 3 ] . While this has improved efficiency, an open gap remains: current systems are often limited to pre-programmed sample transportation. This hinders fully autonomous, end-to-end operation in scientific labs, where robots need to adapt to novel protocols and handle heterogeneous materials, which is a common challenge in early-stage discovery. Efficient and reliable handling of materials, from fine powders to granular crystals, is essential for chemistry lab automation. To date, advanced robot chemists primarily use position controllers [ 20 , 7 ] even for contact-rich manipulation tasks. These controllers are ill-suited for tasks requiring compliance and adaptation to variable forces. Impedance controllers [ 10 ] , in contrast, provide the necessary force control and adaptability to manage unpredictable interactions, making them ideal for handling heterogeneous materials. These controllers are especially valuable for tasks like sample scraping [ 20 ] in automated solid-state materials chemistry [ 14 ] and drug development. This process is essential for recovering expensive molecules and requires a precise vertical motion to retrieve powdered or crystalline samples from vial walls. The task demands adaptive force control to handle variable material properties e.g. , adhesive and cohesive, as well as visual feedback to gauge how much material remains. The challenge is further compounded by the tools themselves. Standard laboratory spatulas are often compliant and bendable, which complicates relying solely on reactive force control. The force measured at the robot’s wrist does not directly correspond to the force at the tool tip due to tool deformation. Prior works on automating this process [ 20 ] have not addressed these requirements. We address these challenges with a novel adaptive control framework that combines a low-level Cartesian Impedance Controller (CIC) with a high-level Reinforcement Learning (RL) agent (Fig. 1 ). The RL agent learns a perception-based, feedforward wrench command that is robust to both unmodelled tool and material dynamics. The RL agent learns an optimal force policy in real-time by interacting with the environment, allowing it to adapt to varying material properties like hardness and adhesion. To ensure the robot applies force only where it is needed, we integrate visual (RGB-D) feedback, which informs the agent of the sample’s location. By isolating force learning from high-dimensional joint control, our method simplifies the learning process and enhances sim-to-real transfer. This framework paves the way for developing adaptive, force-aware robotic chemists for laboratory automation, enabling previously infeasible operations such as scraping heterogeneous material samples. In summary, the contributions of this work are: (1) A novel adaptive control framework for combining a low-level Cartesian impedance controller with a high-level RL agent for learning the desired interaction wrenches to perform material removal from vial walls during autonomous sample scraping. This work presents a perception-driven, ‘force-aware’ approach to the challenging problem of in-vial manipulation with heterogeneous materials; (2) A multi-stage perception pipeline to autonomously localise the vial and accurately detect the sample material for the scraping task. This enables the robot to execute the task without a priori knowing the material’s position or distribution; (3) A demonstration and empirical evaluation of the learned robot behaviour both in simulation and in a real-world chemistry lab across different materials. II Related Work II-A Robot Skill Learning for Laboratory Automation Self-driving laboratories are evolving from simple open-loop automation towards adaptive, modular, and reconfigurable robotic systems [ 27 ] . While perception-guided long-horizon laboratory skills have been proposed for some tasks [ 7 ] , these often rely on basic manipulation skills that primarily involve the pick-and-place of laboratory objects, without handling materials. While these systems are adept at performing tedious and repetitive tasks, they lack the dexterity and skill of human chemists. Current approaches for acquiring these laboratory skills ( i.e. , low-level behaviours) typically fall into two categories: either they are learned within physics-based digital twins using reinforcement learning (RL) and then transferred to the real setup [ 20 , 18 ] , or they use multi-modal