1 Introduction
Empowering robots with a capability to autonomously manipulate liquids will lead to impact across sectors such as engineering and the service industry. Recent approaches have focused on learning how to pour by reasoning using simulations of the liquid Schenck2017 ; CoRL17 or optimization based on parametric assumptions (parabolic trajectory) of the liquid Pan2016_heuristic . The physical parameters of the setup play an important role in both approaches. This includes the shapes of the pouring and receiving containers, intrinsic and extrinsic properties of the liquid, etc. Prior works have focused on inferring specific subsets of these parameters via sampling Schenck2017 or by feedback in closedloop Schenck2017_2 .
A general class of methods, popularly known as intuitive physics battaglia2013simulation ; bates2015humans ; wu2016physics , argues that coarse representations of physical processes are sufficient for many prediction tasks. Inspired by this approach, LopezGuevara et al CoRL17 used an approximate (but realtime) fluid simulator NVIDIA Flex to represent the behavior of liquids. Although this enabled fast prediction, their inference problem involves mapping real world observations to the parameter space of the approximate simulator via a cumbersome calibration step involving pouring.
In this paper, we focus on the problem of inferring the behavior of different types of liquids using simple training interactions and their observed effects. Rather than learning physical properties, we parameterize liquids according to inputs specified to NVIDIA Flex. The learning algorithm searches this parameter space online. For this, the robot stirs the liquid using a motion pattern and seeks simulation parameters that match the inclination of the stick in simulation against its observed values. Finally, we use the inferred parameters to predict the optimized pouring action for a given liquid and verify that it reduces spillage. The highlevel contributions of this paper, in the context of pouring liquids, are that we: (1) decouple actions performed during training from the goal; (2) propose an online, autonomous calibration action; (3) achieve adaptability to different liquids.
2 Related Work and Contribution
Estimation of physical properties: A few approaches focus on estimating physical parameters such as volume Do2016 and viscosity elbrechter2015discriminating ; saal2010active , These methods exploit special measurement equipment such as RGBD cameras or tactile sensors for parameter estimation. In our context, knowledge of the physical parameters would only be useful if a highfidelity simulation is used to optimize decisionmaking during manipulation of fluids. To remain practical, it is necessary to resort to approximate simulators which typically face model mismatch since their input parameters do not coincide exactly with physical attributes such as viscosity. Different approaches have been proposed to learn simulator parameters from data wu2016physics ; Lintusaari2017 ; Gutmann2016a . Different to (wu2016physics, ), we do not assume a Gaussian likelihood of the real observations given the simulated ones. Rather, we learn a model of the discrepancy between real and simulated data and use it to accelerate the search using Bayesian OptimizationGutmann2016a .
Robots interacting with fluids: Existing methods that can reason about fluids, use only simulations Yamaguchi2016 ; Kunze2015 ; Pan2017 or a combination of simulation and real observations Yamaguchi2016stereo . The latter categories of approaches suffer from the problem that approximate simulations deviate over time from reality complicating the envisioned effects of robotic manipulations Kunze2015 . An interesting solution, proposes to use simulations in closedloop Schenck2017_2
, by periodically projecting the simulated particles onto the real liquid tracked in imagespace using thermal imaging. There have been a few solutions on the use of supervised learning
Pan2017 ; Sermanet2017 and Bayesian Optimization CoRL17 for pouring liquids.Summary: We are inspired to combine promising directions of recent work that use supervision Sermanet2017 for learning to pour from, say video of pouring actions. Simultaneously, we retain the benefits of using approximate simulation Schenck2017 ; CoRL17 since it allows a generalization to a variety of manipulations. We decouple the training task from the manipulation so that it lends itself to automation and is less messy for tasks such as pouring liquids. Finally, we perform parameter estimation in the space of inputs of the approximate simulator rather than physical units. This enables us to use these parameters directly for predictive tasks by supplying them to the simulator during test execution.
3 Problem Definition
Let denote actions performed in training(“stirring”). Let define the parameters controlling the behaviour of the liquid in the simulationbased model. For each action executed by the robot, let the observable at time be , the inclination of the stick used for stirring. When the same action is executed by the simulator using the parameter , let the resulting inclination be . We define the observed discrepancy (for stirring) over the duration of action as
Let be a pouring action and let denote the corresponding spillage (as a percentage of the poured liquid) observed when is executed.
Here, we analyze the problem of inferring parameters of the liquid, given a space of training (stirring) actions that are different from the space of goal (pouring) actions . We quantify the suitability of by measuring the percentage of liquid spilled while performing optimized oneshot pouring using obtained from
Assumptions: We assume that the shapes (geometry) of the containers are available, or can be estimated using sensors. Also, we rely on the robot’s estimation of its end effector pose, to synchronise simulation with reality.
Inference: Given a specific , say stirring using a fixed motion pattern, the goal of the inference step is to estimate the best in simulation such that the discrepancy is minimal. At each iteration , an action is executed by the robot and in simulation using a hypothesized parameter . The resulting discrepancy , calculated using Eq. 3, together with the parameter are provided to a Bayesian Optimizer that learns a regression of over using a Gaussian process.
An approximation of the likelihood Gutmann2016a
can be computed using the cdf of the standard Normal distribution
and a threshold as (visualized in Fig. 2Middle):Evaluation: We quantitatively evaluate the suitability of for the problem by measuring percentage spillage using an optimized action . This is due to the lack of an existing ground truth of the parameters in the simulator given its approximate nature. Since our contribution concerns the training task, we use a pouring strategy exactly as proscribed by previous work CoRL17 . They use a simulator to identify
, by defining the loss function to be the ratio of the spilled particles to the total number of particles simulated. The
iteration of their method therefore involves executing the simulator with action and . The minimization results in after a finite number of iterations ( in our case). Finally, we execute using the robot and measure the percentage of liquid spilled.4 Experiments and Results
Experimental setup
Stirring: For all our experiments, we used a UR10 robot equipped with a gripper holding a stick so that it is free to pivot at the gripping point. Before stirring begins, the stick is vertical and partly submerged in the liquid. The motion of the end effector is limited to a plane parallel to the ground plane. Due to this motion, and the the resistance encountered by the stick due to the liquid, at any instant , the stick might deviate from its vertical position to . The inclination is intricately dependant on the velocity of the end effector and the physical properties of the liquid and the stick.
is estimated using simple computer vision on the video feed from 2 webcams with image planes orthogonal to
. The position of the end effector of the robot is queried at Hz and supplied to the simulator which replicates the executed action. The inclination produced in simulation at instant is recorded as . The space of stirring actions is discrete and determined by the stirring pattern. In this work we used a cyclic sequence, that visually follows an point star with .Pouring: We replicate the oneshot pouring solution in existing work CoRL17 . For completeness, we review their method here using our notation. The space of pouring actions is two dimensional and continuous. The 2D space is parameterized by a constant angular velocity and the relative distance between source and target containers . After iterations of the optimizer over given , the robot obtains an estimate for the optimal pouring action , which it then executes. We measure the percentage of liquid spilled by the robot over 5 repetitions of the above experiment.
Discussion
Learning by stirring vs learning by pouring: We compared the percentage spillage achieved by our algorithm which learns by stirring against the method proposed in CoRL17 which calibrates by pouring using a training cup. Although it would seem intuitive that applying the same task to train must result in lower spillage under test conditions, our results indicate the contrary. Fig. 2Rightup plots vs , where is the number of iterations of the B.O. used to estimate . Using our stirring approach, the spillage is less than even with only 10 iterations, under half the corresponding figure when the robot was trained with pouring. At iterations, our approach almost achieves zeros spillage (which is lower than learning from pouring at iterations).
Pouring other liquids: We observed a similar trend across three different liquids (Fig. 2Rightdown): as is increased, the spillage reduces. However, the degree of spillage is significantly higher for more glycerin and gel. On further investigation of the video and the simulator, we realized that the excessive spillage for glycerin is due to the unusually high adhesive effect that makes glycerin stick to the pouring container. Unfortunately, this adhesive behaviour cannot be modelled by the simulator on a particleparticle interaction. We conclude that the choice of the approximate simulator, combined with potentially different behavior across training and pouring actions might be a source of error during spillage. However, the capability to infer parameters within a limited gamut of expressibility is still a valuable addition to the toolkits proposed by existing methods.
5 Conclusion
We have presented the first supervised learning algorithm for robotic manipulation of liquids that decouples the training action (stirring) from the final task (pouring) while adapting to liquids with widely different properties. Learning by stirring is preferable to learning by pouring because it is easy to automate, it is time efficient and avoids the mess involved due to spillage. We demonstrated that stirring leads to reduced spillage for water compared to state of the art and also presented results for adapting the pouring to other liquids. We discussed the several design decisions involved, along with quantitative justification and recommendations for prospective usecases.
References
 [1] Christopher Bates, Peter Battaglia, Ilker Yildirim, and Joshua B Tenenbaum. Humans predict liquid dynamics using probabilistic simulation. In CogSci, 2015.

[2]
Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum.
Simulation as an engine of physical scene understanding.
Proceedings of the National Academy of Sciences, 110(45):18327–18332, 2013.  [3] Chau Do, Tobias Schubert, and Wolfram Burgard. A Probabilistic Approach to Liquid Level Detection in Cups Using an RGBD Camera. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2016.
 [4] Christof Elbrechter, Jonathan Maycock, Robert Haschke, and Helge Ritter. Discriminating liquids using a robotic kitchen assistant. In Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on, pages 703–708. IEEE, 2015.

[5]
M.U. Gutmann and J Corander.
Bayesian optimization for likelihoodfree inference of
simulatorbased statistical models.
Journal of Machine Learning Research
, 17(125):1–47, 2016.  [6] Lars Kunze and Michael Beetz. Envisioning the qualitative effects of robot manipulation actions using simulationbased projections. Artificial Intelligence, jan 2015.
 [7] J. Lintusaari, M.U. Gutmann, R. Dutta, S. Kaski, and J. Corander. Fundamentals and recent developments in approximate Bayesian computation. Systematic Biology, 66(1):e66–e82, January 2017.
 [8] Tatiana LopezGuevara, Nicholas K. Taylor, Michael U. Gutmann, Subramanian Ramamoorthy, and Kartic Subr. Adaptable pouring: Teaching robots not to spill using fast but approximate fluid simulation. In 1st Annual Conference on Robot Learning, CoRL 2017, Mountain View, California, USA, November 1315, 2017, Proceedings, pages 77–86, 2017.
 [9] Zherong Pan and Dinesh Manocha. Motion Planning for Fluid Manipulation using Simplified Dynamics. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), volume 0, 2016.
 [10] Zherong Pan and Dinesh Manocha. Feedback Motion Planning for Liquid Pouring Using Supervised Learning. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017.
 [11] Hannes Saal, JoAnne Ting, and Sethu Vijayakumar. Active sequential learning with tactile feedback. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pages 677–684, 2010.
 [12] Connor Schenck and Dieter Fox. Reasoning About Liquids via ClosedLoop Simulation. In Robotics: Science and Systems (RSS), 2017.
 [13] Connor Schenck and Dieter Fox. Visual ClosedLoop Control for Pouring Liquids. In International Conference on Experimental Robotics (ICRA), 2017.

[14]
Pierre Sermanet, Corey Lynch, Jasmine Hsu, and Sergey Levine.
TimeContrastive Networks: SelfSupervised Learning from Multiview
Observation.
IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops
, 2017July:486–487, 2017.  [15] Jiajun Wu, Joseph J Lim, Hongyi Zhang, Joshua B Tenenbaum, and William T Freeman. Physics 101: Learning physical object properties from unlabeled videos. In BMVC, volume 2, page 7, 2016.
 [16] Akihiko Yamaguchi and Christopher G Atkeson. Differential Dynamic Programming for GraphStructured Dynamical Systems : Generalization of Pouring Behavior with Different Skills. In IEEERAS International Conference on Humanoid Robots, number 2, 2016.
 [17] Akihiko Yamaguchi and Christopher G Atkeson. Stereo Vision of Liquid and Particle Flow for Robot Pouring. In IEEERAS International Conference on Humanoid Robots, number c, 2016.