Helping robots handle liquids – Robohub

Imagine that you are enjoying a picnic by the river on a windy day. A gust of wind accidentally catches your paper napkin and lands on the surface of the water, quickly moving away from you. You grab the nearest stick and carefully stir the water to scoop it up, creating a series of tiny waves. This wave eventually pushed the napkin back to shore, so you took it. In this scenario, the water acts as a force transmission medium, allowing you to manipulate the position of the napkin without direct contact.

Humans regularly engage with various types of fluids in their daily lives, but doing so has become a difficult and elusive goal for today’s robotic systems. Gave you a latte? Robots can do that. For? It will need a little more nuance.

FluidLab, a new simulation tool from researchers at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), enhances robotic learning for complex fluid manipulation tasks such as making latte art, ice cream and even manipulating air. The virtual environment offers a versatile set of complex fluid handling challenges, involving both solids and liquids, and many liquids together. FluidLab supports solid, liquid, and gas modeling, including elastic, plastic, rigid bodies, Newtonian and non-Newtonian fluids, and smoke and air.

The heart of FluidLab lies in FluidEngine, an easy-to-use physics simulator capable of calculating and simulating various materials and their interactions seamlessly, while leveraging the power of the graphics processing unit (GPU) for faster processing. The engine is “different”, meaning the simulator can incorporate knowledge of physics to model a more realistic physical world, leading to more efficient learning and planning of robotic tasks. In contrast, most existing reinforcement learning methods lack a world model that relies solely on trial and error. This enhanced capability, the researchers said, would allow users to experiment with robotic learning algorithms and toys with the limitations of current robot manipulation capabilities.

To set the stage, the researchers tested the robot’s learning algorithms using FluidLab, discovering and solving unique challenges in fluid systems. By developing intelligent optimization methods, they can effectively transfer this learning from simulations to real-world scenarios.

“Imagine a future where household robots easily help you with everyday tasks, such as making coffee, preparing breakfast or cooking dinner. These tasks involve multiple fluid manipulation challenges. Our benchmarks are the first step towards enabling robots to master these skills, benefiting households and the workplace,” said visiting researcher at MIT CSAIL and research scientist at MIT-IBM Watson AI Lab Chuang Gan, senior author on the new paper on the research. “For example, these robots can reduce wait times and improve the customer experience in a busy coffee shop. FluidEngine is, to our knowledge, the first physics engine of its kind to support multiple materials and couplings while being fully differentiated. With our standard fluid manipulation tasks, researchers can evaluate robot learning algorithms and push the boundaries of what today’s robots are capable of manipulating.”

Liquid fantasy

Over the last few decades, scientists in the robotic manipulation domain have primarily focused on the manipulation of rigid bodies, or on very simple fluid manipulation tasks such as pouring water. Learning manipulation tasks involving real-world fluids can also be an insecure and expensive endeavor.

With fluid manipulation, it’s not always just about the liquid. In many tasks, such as creating the perfect ice cream swirl, mixing solids into a liquid, or paddling through water to move objects, it’s a dance of interaction between liquids and various other ingredients. The simulation environment must support “coupling”, or how two different material properties interact. Fluid manipulation tasks typically require very fine precision, with subtle interactions and material handling, distinguishing them from more straightforward tasks such as pushing blocks or opening bottles.

The FluidLab simulator can quickly calculate how different materials interact with each other.

Assisting the GPU is “Taichi,” a domain-specific language embedded in Python. The system can calculate the gradients (rates of changes in the configuration of the environment with respect to robotic actions) for different types of materials and their interaction (coupling) with each other. This precise information can be used to fine-tune the robot’s movements for better performance. As a result, the simulator enables faster and more efficient solutions, setting it apart from its counterparts.

The 10 tasks the team performed were divided into two categories: using fluids to manipulate hard-to-reach objects, and directly manipulating fluids for specific purposes. Examples include separating liquids, guiding floating objects, transporting goods with jets of water, mixing liquids, making latte art, shaping ice cream, and controlling air circulation.

“Simulators work similarly to how humans use their mental models to predict the consequences of their actions and make decisions when manipulating fluids. This is a significant advantage of our simulator compared to others,” said Carnegie Mellon University PhD student Zhou Xian, another author on the paper. “While other simulators primarily support reinforcement learning, ours supports reinforcement learning and enables more efficient optimization techniques. Leveraging the gradients provided by the simulator supports highly efficient policy search, making it an even more versatile and effective tool.”

The next step

The future for FluidLab looks bright. The current work seeks to transfer the trajectories optimized in the simulation to real-world tasks directly by means of an open loop. For the next step, the team is working to develop closed-loop policies within the simulation that take state or visual observations of the environment as input and perform fluid manipulation tasks in real time, and then transfer the learned policies in real-world scenes. .

This platform is open to the public publicly availableand the researchers hope it will be of use to future studies in developing better methods to solve complex fluid manipulation tasks.

“Humans interact with liquids in everyday tasks, including pouring and mixing liquids (coffee, yogurt, soup, dough), washing and cleaning with water, and more,” said University of Maryland computer science professor Ming Lin, who was not involved. in the research. Work. “For robots to assist humans and serve in a similar capacity for everyday tasks, new techniques for interacting with and handling various fluids with different properties (e.g. material viscosity and density) will be required and remain a major computational challenge for real-time autonomous systems. This work introduces the first comprehensive physics engine, FluidLab, to enable the modeling of various complex fluids and their association with other objects and dynamic systems in the environment. The mathematical formulation of ‘differentiable fluids’ as presented in the paper makes it possible to integrate multipurpose fluid simulation as a network layer in learning-based algorithms and neural network architectures for intelligent systems to operate in real-world applications.

Gan and Xian co-authored the paper with Hsiao-Yu Tung, a postdoc in the MIT Department of Brain and Cognitive Sciences; Antonio Torralba, an MIT professor of electrical engineering and computer science and CSAIL principal investigator; Dartmouth College Assistant Professor Bo Zhu, Columbia University PhD student Zhenjia Xu, and CMU Assistant Professor Katerina Fragkiadaki. The team’s research is supported by the MIT-IBM Watson AI Lab, Sony AI, DARPA Young Investigator Award, NSF CAREER award, AFOSR Young Investigator Award, DARPA Machine Common Sense, and the National Science Foundation.

This research was presented at the International Conference on Learning Representation earlier this month.