pith. sign in

arxiv: 1906.12264 · v1 · pith:QJGQB4AWnew · submitted 2019-06-21 · 💻 cs.RO · cs.AI

Accurate Robotic Pouring for Serving Drinks

Yongqiang Huang , Yu Sun This is my paper

Pith reviewed 2026-05-25 18:31 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords robotic pouringrecurrent neural networkshuman demonstrationsliquid handlingmotor controlpouring accuracygeneralization to new containers
0
0 comments X

The pith

A recurrent neural network trained on human water-pouring demos generates container velocities that let a robot pour with 4 milliliter error from unseen containers.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors train a recurrent neural network solely on recordings of people pouring water and use it to produce angular velocity commands for a robot arm holding a source container. These commands are sent in real time to a motor that tilts the container while liquid flows into a target. When tested on containers never seen during training, the poured volume deviates by as little as 4 milliliters. The same network also controls pouring of oil and syrup, with only modestly higher error for the more viscous liquids.

Core claim

A recurrent neural network trained only on human demonstrations of pouring water outputs angular velocities that, when executed by a motor on a physical robot, achieve pouring errors as low as 4 milliliters from containers not present in the training set and produce comparable accuracy when pouring oil or syrup.

What carries the argument

Recurrent neural network that maps the current pouring state to angular velocity commands at each time step, learned from human demonstration trajectories.

If this is right

  • The learned policy transfers to container geometries absent from training data.
  • The same network produces usable velocities for liquids other than water, including oil and syrup.
  • Real-time execution of the network outputs on physical hardware yields measurable volume accuracy without additional sensors during pouring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The result implies that the network has extracted a general pouring policy rather than memorizing specific container shapes.
  • The approach could be tested on targets that move during pouring or on tasks that require exact target volumes.
  • Adding visual sensing of the liquid level might reduce error further when liquid properties differ markedly from water.

Load-bearing premise

The velocities produced by a network trained only on water demonstrations will remain suitable when the container shape changes or the liquid viscosity changes.

What would settle it

Run the system on a container whose geometry differs sharply from the training set and measure whether the volume error stays near 4 milliliters or rises substantially.

Figures

Figures reproduced from arXiv: 1906.12264 by Yongqiang Huang, Yu Sun.

Figure 1
Figure 1. Figure 1: An illustration of the six input features/dimensions for the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cups for training and evaluation be sequential. At any time step during pouring, the generator should take the current poured volume as input, compare it with the target volume, and adjust the velocity accordingly. We use RNN to model the velocity generator. RNN is a class of neural networks that is designed to process its inputs in order. It feeds its output from one time step into its input to the next t… view at source ↗
read the original abstract

Pouring is the second most frequently executed motion in cooking scenarios. In this work, we present our system of accurate pouring that generates the angular velocities of the source container using recurrent neural networks. We collected demonstrations of human pouring water. We made a physical system on which the velocities of the source container were generated at each time step and executed by a motor. We tested our system on pouring water from containers that are not used for training and achieved an error of as low as 4 milliliters. We also used the system to pour oil and syrup. The accuracy achieved with oil is slightly lower than but comparable with that of water.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript presents a robotic pouring system that uses a recurrent neural network trained exclusively on human demonstrations of pouring water to generate angular-velocity sequences for the source container at each time step. These velocities are executed open-loop by a motor on a physical setup. The central experimental claim is that the system achieves pouring errors as low as 4 ml on water from containers unseen during training and produces accuracy for oil and syrup that is slightly lower but comparable to water.

Significance. If the reported accuracy and cross-liquid generalization hold under rigorous testing, the work would provide a practical demonstration of learning pouring policies from human data that transfer to novel geometries and unmodeled fluid properties without online sensing or liquid-specific conditioning. This could support data-driven approaches to precise liquid manipulation in robotic cooking and serving applications.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'an error of as low as 4 milliliters' for water on unseen containers supplies no supporting statistics (number of trials, mean error, standard deviation, error bars) or baseline comparisons, rendering the quantitative result unverifiable from the given text.
  2. [Abstract] Abstract: the generalization claim for oil and syrup is stated only qualitatively ('slightly lower than but comparable with that of water') with no numerical error values, no discussion of viscosity differences, and no indication of whether the RNN receives liquid-type input or relies solely on water-trained velocities in open-loop execution.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'an error of as low as 4 milliliters' for water on unseen containers supplies no supporting statistics (number of trials, mean error, standard deviation, error bars) or baseline comparisons, rendering the quantitative result unverifiable from the given text.

    Authors: The reported figure of 4 ml represents the lowest error observed across trials with unseen containers. The full manuscript provides the underlying experimental data including trial counts and variability. We will revise the abstract to report the number of trials, mean error, standard deviation, and relevant baseline comparisons to make the claim self-contained and verifiable. revision: yes

  2. Referee: [Abstract] Abstract: the generalization claim for oil and syrup is stated only qualitatively ('slightly lower than but comparable with that of water') with no numerical error values, no discussion of viscosity differences, and no indication of whether the RNN receives liquid-type input or relies solely on water-trained velocities in open-loop execution.

    Authors: The RNN is trained exclusively on water demonstrations and receives no liquid-type input; the identical open-loop velocity commands are applied to oil and syrup. We agree the abstract should be more precise. We will revise it to include the numerical error values obtained for oil and syrup, along with a brief statement on the absence of liquid-specific conditioning and the role of viscosity differences. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results from physical tests on unseen containers

full rationale

The paper describes collecting human water-pouring demonstrations, training an RNN to output angular velocities, and executing them open-loop on a physical motor system. Reported errors (as low as 4 ml for water on unseen containers; comparable for oil/syrup) are presented as direct physical measurements. No equations, fitted parameters, self-citations, or derivations are shown that would reduce any claimed result to its inputs by construction. The central claim rests on experimental generalization rather than any self-referential definition or renaming.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that the learned policy transfers across containers and liquids.

pith-pipeline@v0.9.0 · 5619 in / 999 out tokens · 19901 ms · 2026-05-25T18:31:04.985431+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 2 internal anchors

  1. [1]

    Brandi, O

    [Brandi et al., 2014] S. Brandi, O. Kroemer, and J. Pe- ters. Generalizing pouring actions between objects using warped parameters. In 2014 IEEE-RAS International Con- ference on Humanoid Robots, pages 616–621, Nov

    work page 2014

  2. [2]

    Accurate pouring using model predictive con- trol enabled by recurrent neural network

    [Chen et al., 2019] Tianze Chen, Yongqiang Huang, and Yu Sun. Accurate pouring using model predictive con- trol enabled by recurrent neural network. In Intelligent Robots and Systems (IROS), 2019 IEEE/RSJ International Conference on, pages 1–7. IEEE,

    work page 2019

  3. [3]

    Accurate Pouring with an Autonomous Robot Using an RGB-D Camera

    [Do and Burgard, 2018] Chau Do and Wolfram Burgard. Ac- curate pouring with an autonomous robot using an RGB-D camera. CoRR, abs/1810.03303,

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  4. [4]

    Learning to pour using deep deterministic pol- icy gradients

    [Do et al., 2018] Chau Do, Camilo Gordillo, and Wolfram Burgard. Learning to pour using deep deterministic pol- icy gradients. In 2018 IEEE International Conference on Robotics and Automation (ICRA) ,

    work page 2018

  5. [5]

    Gers, Nicol N

    [Gers et al., 2003] Felix A. Gers, Nicol N. Schraudolph, and J¨urgen Schmidhuber. Learning precise timing with lstm re- current networks. J. Mach. Learn. Res., 3:115–143, March

    work page 2003

  6. [6]

    Generating Sequences With Recurrent Neural Networks

    [Graves, 2013] Alex Graves. Generating sequences with re- current neural networks. CoRR, abs/1308.0850,

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  7. [7]

    Modeling dynamic system by recurrent neural network with state variables

    [Han et al., 2004] Min Han, Zhiwei Shi, and Wei Wang. Modeling dynamic system by recurrent neural network with state variables. In Advances in Neural Networks - ISNN 2004,

    work page 2004

  8. [8]

    Huang and Y

    [Huang and Sun, 2017] Y . Huang and Y . Sun. Learning to pour. In 2017 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS) , pages 7005–7010, Sept

    work page 2017

  9. [9]

    A dataset of daily interactive manipulation

    [Huang and Sun, 2019] Yongqiang Huang and Yu Sun. A dataset of daily interactive manipulation. The International Journal of Robotics Research , 0(0):0278364919849091,

    work page 2019

  10. [10]

    Recent data sets on object ma- nipulation: A survey

    [Huang et al., 2016] Yongqiang Huang, Matteo Bianchi, Mi- nas Liarokapis, and Sun Yu. Recent data sets on object ma- nipulation: A survey. Big Data, 4(4):197–216, December

    work page 2016

  11. [11]

    Paulius, Y

    [Paulius et al., 2016] D. Paulius, Y . Huang, R. Milton, W. D. Buchanan, J. Sam, and Y . Sun. Functional object-oriented network for manipulation learning. In 2016 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pages 2655–2662, Oct

    work page 2016

  12. [12]

    Paulius, A

    [Paulius et al., 2018] D. Paulius, A. B. Jelodar, and Y . Sun. Functional object-oriented network: Construction expan- sion. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 1–7, May

    work page 2018

  13. [13]

    Schenck and D

    [Schenck and Fox, 2017] C. Schenck and D. Fox. Visual closed-loop control for pouring liquids. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2629–2636, May

    work page 2017

  14. [14]

    Trischler and Gabriele M.T

    [Trischler and D’Eleuterio, 2016] Adam P. Trischler and Gabriele M.T. D’Eleuterio. Synthesis of recurrent neural networks for dynamical system simulation. Neural Net- works, 80:67 – 78, 2016

    work page 2016