Tool-Policy Co-Design for Powder Weighing in Laboratory Automation

Nikola Radulov; Xin Yang; Kevin S. Luck; Gabriella Pizzuto

cs.ROSep 30, 2026

Tool-Policy Co-Design for Powder Weighing in Laboratory Automation

Authors: Nikola Radulov, Xin Yang, Kevin S. Luck, Gabriella Pizzuto

Organizations: School of Computer Science & Informatics, University of Liverpool, UK. · Vrije Universiteit Amsterdam, Faculty of Science, Netherlands. · Department of Chemistry, University of Liverpool, UK.

Abstract

Autonomous powder weighing is one of many bottlenecks in laboratory automation due to the complex, non-linear dynamics of heterogeneous materials. Robot chemists performing this task utilise standard tools shaped for the dexterity of human hands, whose fixed geometry sets the dynamics that the control policy needs to regulate. This work introduces a tool-policy co-design framework that concurrently optimises the morphology of a dispensing tool and its control policy for use by robots in chemistry laboratories, formulated as a bi-level optimisation that minimises dispensing error over a target distribution of powder flowabilities. The outer loop varies tool-design parameters such as tool depth, width and rim spike topology using Bayesian optimisation and hyperband, while an inner loop optimises a control policy for each candidate morphology. We also introduce a geometric similarity metric that warm-starts policy training from cached policies of structurally similar designs, exploring 28% more configurations under the same compute budget. The proposed framework is evaluated on a robotic powder weighing task across seven materials with distinct physical dynamics in a flowability-informed robot-material simulation framework. Experimental results demonstrate that our co-designed tool morphology reduces real-world weighing errors by 45% relative to a standard tool, including on previously unseen materials. These results demonstrate our method can adapt both the control policy and the physical tool to the dynamics of the target material, bringing a new paradigm for material manipulation to the field of laboratory automation.

Figures & tables

Fig. 1 : Tool-policy co-design method for powder weighing. We use BOHB to propose tool morphologies $\xi$ . For each candidate $\xi$ , a control policy $\pi_{\xi}$ is trained in a low-fidelity environment and evaluated in a high-fidelity digital twin. The highest-performing design and paired policy are transferred for real-world validation.

Fig. 2 : Tool morphology parameters of the design space; $l$ , $h$ , $w$ represent the length, depth and width of the spoon bowl. $H_{k}$ represents the height of the $k^{th}$ spike.

Panel of Figure 3 — Fig. 3 : The performance of each chosen morphology as the mean final test error and the respective standard deviation, averaged across 3 seeds. The temperature represents the mean final test error ( $\text{\,}\mathrm{m}\mathrm{g}$ ). The top graph ( 3(a) ) shows the raw task error when evaluated on the low-fidelity environment and the bottom graph ( 3(b) ) shows the performance when evaluated on the high-fidelity test environment.

Fig. 4 : Sim-to-real rank shifts across morphologies. Each spoon image shows its baseline simulation rank; blue arrows show shifts for IID powders.

\Block 2-4 Spoon Configuration				Powder Weighing Error (mg)							\Block 3-1In-distribution Error	\Block 3-1Overall Error
				\Block 2-1Sand	\Block 2-1Sugar	\Block 2-1Salt	\Block 2-1Semolina	\Block 2-1Sodium Bicarbonate	\Block 2-1Pectin	\Block 2-1Flour
No.	Depth	Width	Spike Configuration
Standard Tool				$1.49\pm 1.36$	$2.05\pm 1.73$	$1.66\pm 1.19$	$2.52\pm 3.66$	$4.45\pm 5.57$	$\mathbf{3.58\pm 3.64}$	$12.91\pm 13.58$	$1.93\pm 2.16$	$4.09\pm 6.82$
1	1.1	1.1	{1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}	$2.12\pm 4.21$	$4.78\pm 5.02$	$12.56\pm 2.98$	$*$	$14.48\pm 0.23$	$13.15\pm 3.96$	$13.29\pm 4.06$	$8.58\pm 6.36$	$10.74\pm 5.79$
2	1.1	0.7	{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0}	$\mathbf{0.67\pm 0.63}$	$\mathbf{0.53\pm 0.53}$	$1.01\pm 0.54$	$\mathbf{2.47\pm 1.41}$	$3.62\pm 3.45$	$8.21\pm 4.04$	$11.41\pm 8.72$	$\mathbf{1.17\pm 1.13}$	$\mathbf{4.04\pm 5.43}$
3	1.5	1.5	{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0}	$4.38\pm 2.27$	$2.29\pm 1.27$	$1.66\pm 1.49$	$14.48\pm 0.99$	$14.49\pm 0.97$	$*$	$14.66\pm 0.34$	$5.70\pm 5.44$	$9.52\pm 6.06$

TABLE I : Zero-shot transfer results on

w_{target}=15

\text{\,}\mathrm{m}\mathrm{g}

of selected policy-tool pairs, evaluated using the best-performing policy in simulation for each morphology. For each powder, results report the mean absolute error and standard deviation over 10 real-world runs. Sodium bicarbonate, pectin and flour represent OOD materials, i.e. , the policy was not trained on these specific powders or materials with similar flow. Data points marked with

*

represent cases where no material was dispensed.

Panel of Figure 6 — Fig. 5 : Top three morphologies discovered by the BOHB optimiser. The simulated models are shown in the top row, with their corresponding physical 3D-printed prototypes below. The table details the specific width, depth, and spike configurations for each design.

\Block 2-1Configuration	Powder Weighing Error (mg)				\Block 2-1Simulation Error	\Block 2-1Real-world Error
	Sand	Sugar	Salt	Semolina
Config a	$0.80\pm 0.45$	$0.96\pm 0.89$	$1.03\pm 0.89$	$2.58\pm 1.58$	$0.97\pm 0.31$	$1.31\pm 1.17$
Config b	$\mathbf{0.55\pm 0.24}$	$\mathbf{0.71\pm 0.35}$	$\mathbf{0.65\pm 0.73}$	$1.85\pm 1.43$	$0.94\pm 0.09$	$\mathbf{0.91\pm 0.93}$
Config c	$0.76\pm 0.58$	$1.0\pm 0.9$	$1.18\pm 1.44$	$\mathbf{1.52\pm 1.36}$	$\mathbf{0.85\pm 0.17}$	$1.10\pm 1.10$

TABLE II : Simulation and real-world deployment comparison of the best 3 tool configurations found by BOHB, reported from the high-fidelity simulation environment and for IID powders, averaged over 10 real-world runs for

w_{target}=15

\text{\,}\mathrm{m}\mathrm{g}

Fig. 6 : Each point is two-dimensional representation of a spoon configuration. Spoon designs that are located closer together in the 2D space have higher morphological similarity. The colour of the points represents their task error in simulation.

Fig. 7 : BOHB performance with and without morphology similarity filter. The plot shows total unique sampled configurations at $50000$ , $125000$ and $200000$ total episodes used.

\Block 2-1Configuration	Powder Weighing Error (mg)				\Block 2-1Simulation Error	\Block 2-1Real-world Error
	Sand	Sugar	Salt	Semolina
Config A	$0.65\pm 0.54$	$\mathbf{0.47\pm 0.24}$	$0.70\pm 0.36$	$1.82\pm 1.25$	$0.95\pm 0.05$	$0.91\pm 0.87$
Config B	$0.62\pm 0.55$	$0.56\pm 0.23$	$\mathbf{0.41\pm 0.18}$	$\mathbf{1.66\pm 0.92}$	$0.92\pm 0.08$	$\mathbf{0.81\pm 0.73}$
Config C	$\mathbf{0.59\pm 0.27}$	$0.73\pm 0.50$	$0.91\pm 1.33$	$1.97\pm 0.91$	$\mathbf{0.86\pm 0.22}$	$1.05\pm 0.98$

TABLE III : Simulation and real-world deployment comparison of the best 3 tool configurations found by BOHB with similarity, reported from the high-fidelity simulation environment and for IID powders, averaged over 10 real-world runs for

w_{target}=15

\text{\,}\mathrm{m}\mathrm{g}

Morphology	IID Error	Sodium Bicarbonate ( $43^{\circ})$	Pectin ( $46^{\circ}$ )	Flour ( $51^{\circ}$ )	Overall Real-world Error
Standard Tool	$1.93\pm 2.16$	$4.45\pm 5.57$	$3.58\pm 3.64$	$12.91\pm 13.58$	$4.09\pm 6.82$
Config. 2	$1.17\pm 1.13$	$3.62\pm 3.45$	$8.21\pm 4.04$	$11.41\pm 8.72$	$4.04\pm 5.43$
Config. B	$\mathbf{0.81\pm 0.73}$	$\mathbf{1.98\pm 1.53}$	$\mathbf{2.71\pm 2.43}$	$\mathbf{7.69\pm 4.01}$	$\mathbf{2.23\pm 3.00}$

TABLE IV : Zero-shot transfer results on OOD materials with their respective AoR, averaged over 10 runs for

w_{target}=15

\text{\,}\mathrm{m}\mathrm{g}

Explore similar work

Labimus: A Simulation and Benchmark for Humanoid Dexterous Manipulation in Chemical Laboratory
Jun 30, 2026Yuhan Wu, Zhao Jin, Tao Li +9Laboratory Precision
University of Science and Technology of China · 2Beijing Innovation Center of Humanoid Robotics
House of Dextra: Cross-embodied Co-design for Dexterous Hands
Dec 3, 2025Kehlani Fay, Darin Anthony Djapri, Anya Zorin +6Robotic Hands Dexterous Manipulation
University of California, San Diego · University of California, Santa Barbara
Scalable Low-Cost Laboratory Automation: A Digital Twin-Integrated Robotic Platform for Autonomous Liquid Handling (RAINBOT)
Jul 22, 2026Mohamed Rami Ayeche, Souhil Sid, Ahyen Mostofa +3Laboratory Robocasa
Acceleration Consortium, University of Toronto, Toronto, Canada. · AISCIA Informatics, Qatar Science and Technology Park, Doha, Qatar · Hamad Bin Khalifa University (HBKU), Education City, Doha, Qatar.

Jun 30, 2026, cs.ROJ/K move · Enter open · S save

Labimus: A Simulation and Benchmark for Humanoid Dexterous Manipulation in Chemical Laboratory

Yuhan Wu, Zhao Jin, Tao Li, Yuheng Zhang+8

University of Science and Technology of China · 2Beijing Innovation Center of Humanoid Robotics

Laboratory Precision

Open paper AskPDF arXiv

Laboratory automation has made remarkable progress through robotic platforms and AI-driven scientific reasoning. However, many laboratory operations (e.g., solid--solid transfer) remain inherently dynamic and require real-time adaptation to different materials and experimental conditions. Such precision-critical manipulations are difficult to standardize, motivating the use of humanoid robots with dexterous hands. Despite this opportunity, no existing benchmark evaluates humanoid manipulation in precision-critical laboratory environments. We present Labimus, to our knowledge, the first benchmark for humanoid dexterous manipulation in organic chemistry laboratories. Labimus reconstructs over 30 functionally faithful assets from real organic chemistry workstations through real-to-sim modeling, collectively covering the core operations of routine organic chemistry experiments. The benchmark integrates articulated laboratory instruments, particle-based powder physics, and closed-loop instrument readouts, enabling a complete manipulation-to-measurement pipeline. It further defines six atomic operations and a seven-step solid-weighing workflow derived from real laboratory standard operating procedures. We introduce a precision-aware evaluation protocol designed to jointly measure task completion, experimental precision, and long-horizon execution. We benchmark three representative policies under procedural layouts and environmental perturbations. Results reveal a precision gap: policies that successfully complete laboratory tasks can still fail to satisfy the quantitative tolerances required by experimental protocols. Our benchmark exposes a fundamental disconnect between task completion and experimental validity, providing a new testbed for developing reliable humanoid robots for scientific laboratories.

Abstract

Explore similar work

Labimus: A Simulation and Benchmark for Humanoid Dexterous Manipulation in Chemical Laboratory

House of Dextra: Cross-embodied Co-design for Dexterous Hands

Scalable Low-Cost Laboratory Automation: A Digital Twin-Integrated Robotic Platform for Autonomous Liquid Handling (RAINBOT)