CMU publishes new dexterous robot algorithm that accurately learns how to operate everyday furniture

Most of the furniture that people come into contact with in daily life are "articulated objects", such as drawers with pull-out rails, doors with vertical rotation axes, doors with horizontal rotation Shaft oven, because the main parts of these objects are connected by various joints.

Due to the existence of these joints, the various parts of the parts of the connected object are kinematically constrained by the joints, so these parts have only one degree of freedom (1 DoF). These items are everywhere in our lives, especially in our daily homes. They are an important part of our daily lives. We as humans see that no matter what kind of furniture we have, we can quickly figure out how to manipulate and control it. It's as if we know how every joint of these objects moves.

So can robots predict how furniture will move like humans? This kind of predictive ability is hard to come by, and if robots could learn this ability, it would be a huge boost for home robots.

Recently, two students in the R-PAD laboratory of Professor David Held of the CMU School of Robotics, Ben Eisner and Harry Zhang, have made breakthroughs in manipulating complex joint objects and launched a 3D-based FlowBot 3D for Neural Networks, an algorithm that effectively expresses and predicts the motion trajectories of parts of articulated objects, such as everyday furniture. The algorithm contains two parts.

The first part is the perception part, which uses a 3D deep neural network to predict the three-dimensional instantaneous motion trajectory from the point cloud data of the manipulated furniture object ( 3D Articulated Flow).

The second part of the algorithm is the policy part, which uses the predicted 3D Articulated Flow to select the robot's next action. Both are fully learned in the simulator and can be implemented directly in the real world without retraining or tuning. With the help of the FlowBot 3D algorithm, the robot can manipulate articulated objects such as everyday furniture at will, just like humans.

##This paper is currently the world’s top robotics conference Robotics Science and Systems (RSS) 2022 best paper candidates (top 3%), and will be exhibited in New York, USA in July, competing with 7 other outstanding articles for the honor of the best paper.

• Paper address: https://arxiv.org/pdf/2205.04382.pdf
• Project homepage: https:// sites.google.com/view/articulated-flowbot-3d

FlowBot 3D relies solely on the simulator and performs supervised learning on simulated data to learn articulated objects such as everyday furniture Instantaneous motion trajectory of the part (3D Articulated Flow). 3D Articulated Flow is a visual point cloud trajectory representation method that can greatly simplify the complexity of the robot's next strategy and improve generalization and efficiency. The robot can complete the task of manipulating joint objects by simply following this instantaneous trajectory and re-predicting this trajectory in a closed loop.

Previously, the conventional method in academia for manipulating joint objects such as furniture was to calculate the movement direction of the part through the geometric characteristics of the manipulated object (such as the position and direction of the connected parts). , or by imitating expert strategies (usually from humans) to learn the operation of specific objects to complete complex actions of joint object manipulation. These traditional methods in academia do not have good generalization, and the efficiency of utilizing data is low. Training requires the collection of a large amount of human demonstration data. Unlike these, FlowBot 3D is the first purely simulator-based learning that does not require humans to provide any demonstration data, and the algorithm allows the robot to calculate the optimal object manipulation path by learning the instantaneous motion trajectory of each part, so the The algorithm has great generalizability. It is this feature that allows FlowBot 3D to generalize to objects invisible during simulator training, successfully manipulating real, everyday furniture items directly in the real world.

The following animations demonstrate the manipulation process of FlowBot 3D. On the left is the manipulated video, and on the right is the predicted instantaneous motion trajectory of the point cloud 3D Articulated Flow. The FlowBot 3D algorithm first enables the robot to identify which part on an object can be manipulated and then predict the direction of movement of that part.

Open the refrigerator door:

Open the toilet seat:

##Open the drawer:

##The reviewer of this paper said: Overall, this paper It is a considerable contribution to robot control science.

So, how does FlowBot 3D learn this skill?

When humans see a new furniture item, such as a door, we know that the door rotates through a door axis, and we know the constraints of the door axis. This door can only rotate in one direction, so we can follow the direction imagined in our minds to open the door. Therefore, if you want a robot to be truly dexterous and effective in predicting the manipulation methods and motion trajectories of joint objects such as furniture, an effective way is to let the robot understand the kinematic constraints of these parts, so that it can predict the movement of these objects. trajectory.

The specific method of FlowBot 3D is not complicated and relies only on the simulator without the need for complicated real human data. In addition, another benefit of the simulator is that in the simulator, the 3D data files (URDF) of these household objects contain the kinematic constraints of each part and the specific parameters of the constraints, so the motion trajectory of each part is in the simulator. can be calculated accurately.

Two modules for FlowBot 3D.

During simulator training, the robot observes the three-dimensional point cloud data of the manipulated object as input data to the robot vision module. The vision module (perception module) uses PointNet to predict the 3D articulated flow of the instantaneous motion trajectory of each point in the input point cloud under the action of external force (for example, after the drawer is opened 1cm, the door opens outward 5 degrees), using the three-dimensional coordinate vector expressed in poor form. The actual data of this motion trajectory can be accurately calculated through forward kinematics. By subtracting the current three-dimensional vector coordinate from the next three-dimensional vector coordinate, the motion trajectory of the manipulated object part can be obtained. Therefore, during training, only the L2 loss of the predicted 3D Articulated flow needs to be minimized for supervised learning.

In this picture, the blue points are the observed point cloud data, and the red arrows represent the predicted facade. Motion trajectory 3D Articulated Flow.

By learning in this way, FlowBot 3D can learn the movement direction of each part under kinematic constraints and the situation where each point on the part is subject to the same force. The relative speed and relative direction of motion (velocity). Common household joint items are prismatic and revolute. For twitch parts, such as drawers, the movement direction and speed of each point on the drawer surface are the same when receiving the same external force. For rotating parts, such as doors, the direction of movement of each point on the door surface is the same when receiving the same external force, but the speed increases the further away from the rotation axis. The researchers used the physical laws in robotics (screw theory) to prove that the longest 3D Articulated Flow can maximize the acceleration of the object. According to Newton's second law, this strategy is the optimal solution.

#Based on theoretical basis, in actual operation , what the robot needs to do is to predict the movement trajectory of each point through the vision module of FlowBot 3D. In each point trajectory, find the point corresponding to the longest 3D articulated flow direction as the manipulation point, and predict this manipulation in a closed loop The movement trajectory of the point. If the selected manipulation point cannot be successfully grasped (for example, the surface does not meet the grasping conditions of the robot hand), then FlowBot 3D will select the point with the second longest length that meets the grasping conditions.

In addition, due to the characteristics of PointNet, FlowBot 3D predicts the motion trajectory of each point and does not rely on the geometric characteristics of the object itself. It has a strong influence on the possible occlusion of the object by the robot. robustness. In addition, because this algorithm is closed-loop, the robot can correct its possible errors in the next step of prediction.

FlowBot 3D’s performance in the real worldFlowBot 3D has the ability to overcome generalization challenges in the real world. The design concept of FlowBot 3D is that as long as it can accurately predict the movement trajectory of the manipulated object 3D articulated flow, then the next step is to follow this trajectory to complete the task.

Another important point is that FlowBot 3D uses a single training model to manipulate multiple categories of items, including categories that have not been seen in training. And in the real world, the robot only needs to use the model obtained through this pure simulator training to manipulate a variety of real objects. Therefore, in the real world, since the kinematic constraints of household objects are overwhelmingly the same as in the simulator, FlowBot 3D can be directly generalized to the real world.

Household items used by FlowBot3D in real-world experiments (including trash cans, refrigerators, toilet seats, boxes, safes, etc.

In the simulator, the robot is trained using some categories of household items, including staplers, bins, drawers, windows, refrigerators, etc. In the simulator and real-world tests, test data New objects from training categories and categories not seen during training.

#FlowBot 3D manipulation tasks in the simulator.

In comparison, imitation-based simulations commonly seen in academia The learning method requires manual guidance to learn how to manipulate new objects, making it unrealistic for these robots to be implemented in the real world, especially in home robot scenarios. In addition, 3D point cloud data is stronger than the 2D RGB data used by other methods. Because point cloud allows the robot to understand each joint and the relationship between joints, it can understand and predict the movement trajectory of parts at a higher level, greatly enhancing generalization.

Experimental results show that FlowBot 3D can achieve a distance of less than 10% to "full open" when operating most objects (whether they are categories seen or not seen during training), and Success Ridge can reach More than 90. In comparison, other methods based on imitation learning (DAgger) or reinforcement learning (SAC) are far behind and lack generalization.

In short, FlowBot 3D is a job with great potential. It can be deployed efficiently in the real world without the need for fine-tuning. This work also shows that advances in computer vision can change the field of robotics, especially the visual expression of motion trajectories called 3D articulated flow, which can be applied to multiple tasks to simplify the robot strategy selection and decision-making process. With this generalizable expression, simulator learning methods will have the potential to be directly deployed in the real world, which will greatly reduce the cost of future home robot training and learning.

FlowBot 3D’s next planCurrently, the research team is trying to apply the flow understanding and prediction method to objects other than joint objects, such as how to use flow Predict object trajectories with 6 degrees of freedom. At the same time, the author is trying to use flow as a general visual expression to apply it to other robot learning tasks, such as reinforcement learning, thereby increasing learning efficiency, robustness, and generalizability.

Associate Professor David Held’s homepage: https://davheld.github.io/Ben Eisner’s homepage: https://beisner.me/Harry Zhang’s homepage: https:// harryzhangog.github.io/

The above is the detailed content of CMU publishes new dexterous robot algorithm that accurately learns how to operate everyday furniture. For more information, please follow other related articles on the PHP Chinese website!