One-Shot Learning for Double-Shot Latte Art: Adapting Cobot Pouring Skills with Dynamic Movement Primitives

Zain Khan; Victor Hernandez Moreno; Dinh Tung Le

cs.ROOct 12, 2026

One-Shot Learning for Double-Shot Latte Art: Adapting Cobot Pouring Skills with Dynamic Movement Primitives

Authors: Zain Khan, Victor Hernandez Moreno, Dinh Tung Le

Organizations: UTS Robotics Institute, School of Electrical, Mechanical and Biomedical Engineering, University of Technology Sydney, Australia

Abstract

Latte art is a complex task requiring dexterity and precision, making it a challenging and commercially relevant task. A one-shot learning system is employed in which a single barista demonstration is segmented into five functionally distinct phases, each encoded as an independent Cartesian Dynamic Movement Primitive (CDMP) with its own parameters. Against a single-CDMP baseline, per-phase encoding reduces transition endpoint error by 3.2-4.5 times and art-stroke shape error by 2.7 times, reproducing jug tilt to below 1 degree RMS in the pouring phases. The learnt pour scales from a 280mL to a 400mL cup through a single spatial offset and a scaled recipe, with no re-demonstration. The phase-based CDMP approach generalises beyond latte art to any demonstration task whose phases have distinct spatial or temporal requirements whilst demonstrating the ability to adapt reproductions to suit different task requirements.

Figures & tables

Figure 1: Robot pour in motion with highlight results

#	Phase	Type
1	base_to_mix	transition
2	mixing	pouring
3	mix_to_art	transition
4	art	pouring
5	return_base	transition

Table 1: Phase decomposition of a latte art demonstration

Figure 2: Phase decomposition shown across each layer. Each column is one phase; rows show the human demonstration, the learnt primitives, and the robot reproduction.

	$280\text{\,}\mathrm{ml}$	$400\text{\,}\mathrm{ml}$
Espresso dose	$21\text{\,}\mathrm{g}$ (single)	$42\text{\,}\mathrm{g}$ (double)
Milk volume	$250\text{\,}\mathrm{ml}$	$350\text{\,}\mathrm{ml}$
Rim diameter	$80\text{\,}\mathrm{mm}$	$90\text{\,}\mathrm{mm}$

Table 2: Beverage parameters for the two cup sizes.

Phase	$\alpha_{s}$	$K$	$\rho$	$\tau$
base_to_mix	7.0	500	5.0	1.0
mixing	7.0	800	10.0	1.0
mix_to_art	7.0	700	6.0	4.0
art	7.0	800	8.0	1.0
return_base	7.0	500	5.0	1.0

Table 3: Per-phase CDMP parameters: temporal scaling

\tau

, basis-function density

\rho

, stiffness

K

(

D=2\sqrt{K}

Panel of Figure 3 — Figure 3: Top-down trajectories of the mix_to_art transition for (a) the one-shot and (b) the phased reproduction, showing the human demonstration, the CDMP rollout and the executed robot motion.

			Phase Evaluation Metric				Tilt RMS ( $\mathrm{\SIUnitSymbolDegree}$ )		Height RMS ( $\mathrm{mm}$ )
Phase	Type	Metric	E1	E2	$\times$	Unit	E1	E2	E1	E2
base_to_mix	transition	endpoint error	30.1	9.5	3.2	mm	5.05	1.03	22.1	6.3
mix_to_art	transition	endpoint error	14.9	3.3	4.5	mm	9.51	0.61	60.8	18.7
return_base	transition	endpoint error	38.3	11.9	3.2	mm	9.35	3.54	40.7	12.4
mixing	pouring	coverage ratio	1.18	1.17	1.0	—	4.35	0.88	18.0	15.9
art	pouring	shape RMS	6.35	2.38	2.7	mm	3.15	0.75	11.3	16.5

Table 4: Per-phase comparison of one-shot (E1) and phased (E2) reproduction against the human demonstration. Phases are grouped by type.

\times

denotes the improvement factor E1/E2; lower is better except for coverage ratio, where

1.00

is ideal.

Panel of Figure 4 — Figure 4: Jug tilt relative to ground plane over the art phase for (a) the one-shot and (b) the phased reproduction.

Panel of Figure 5 — Figure 5: Top-down photographs of the produced latte art. E1 is omitted as it was not poured with milk (Section 5.3 ).

Robot Skill Learning

Robot skill learning studies how robots acquire, adapt, and combine reusable behaviors from demonstrations, interaction, and other data.

See all 320 papers

FlowVLA: Learning Flow-Aware Physical Completion for Quantitative Robotic Pouring
Oct 21, 2026Liwen Yang, Yutian Zhang, Jiaqi Zhai +10Robotic Manipulation VLMs for Robotics
Zhejiang University · DEEP Robotics · Shanghai Artificial Intelligence Laboratory +2
Jointly Learning Predicates and Actions Enables Zero-Shot Skill Composition
May 20, 2026Benedict Quartey, Sebastian Castro, Eric Rosen +3Robot Skill Learning Generalization in Robotic Manipulation
Brown University · Robotics & AI Institute
IMPASTO: Integrating Model-Based Planning with Learned Dynamics Models for Robotic Oil Painting Reproduction
Mar 31, 2026Yingke Wang, Hao Li, Yifeng Zhu +6Model-Based Planning
Stanford University. · The University of Texas at Austin. · University of California, Berkeley. +1

Oct 21, 2026J/K move · Enter open · S save

FlowVLA: Learning Flow-Aware Physical Completion for Quantitative Robotic Pouring

Liwen Yang, Yutian Zhang, Jiaqi Zhai, Li Yang+9

Zhejiang University · DEEP Robotics · Shanghai Artificial Intelligence Laboratory+2

Robotic Manipulation VLMs for Robotics Language-Conditioned Robot Manipulation

Open paper AskPDF arXiv

During steady pouring, images and robot pose may remain nearly unchanged while transferred mass accumulates, making the return decision ambiguous for a whole-task vision-language-action (VLA) policy. We present FlowVLA, which separates motor generation from physical-completion estimation. A shared

π_{0.5}

action expert learns approach, pour, and return with boundary-aware supervision. A target-independent estimator fuses visual, joint, wrist-wrench, and recurrent force-belief histories to maintain an estimate of transferred mass. The controller initiates return when this estimate plus predicted residual outflow reaches the requested mass. In coffee-pot/water experiments, FlowVLA achieves 13.1 g goal-balanced mean absolute error (MacroMAE), lower than all tested VLA baselines. Across all attempts, 53.7% both complete the task and finish within +/-10% of the requested mass. The same model interpolates to held-out 150/250-g targets without retraining. Independently trained models further evaluate the formulation in three additional container-fluid settings.

Abstract

Robot Skill Learning

Explore similar work

FlowVLA: Learning Flow-Aware Physical Completion for Quantitative Robotic Pouring

Jointly Learning Predicates and Actions Enables Zero-Shot Skill Composition

IMPASTO: Integrating Model-Based Planning with Learned Dynamics Models for Robotic Oil Painting Reproduction