Automatic learning of pushing strategy for delivery of irregular-shaped objects
Summary (2 min read)
Introduction
- The use of mobile robots has increased greatly over the past decade as they can be found in practical use in factories, hospitals, and people’s homes.
- This paper introduces a learning-based approach for pushing objects of any irregular shape (Fig. 1).
- The strength of their approach is that the authors can use the same algorithm for objects of different shapes and sizes.
- In addition, the data implicitly stores relevant physical parameters such as the weight distribution and the friction between the robot, object, and pushing surface.
- For the second problem, the authors use a non-parametric regression technique to decide where to move the robot and which direction it should push in, given the final goal and the robot/object relative locations.
A. Problem Definition
- There are two problems that the authors solve to achieve their pushing task.
- The authors assume that their system can continuously track the global position/orientation of the robot, object, and goal.
- The first problem is to collect data on the object’s movement based on the robot’s position and push direction.
- The authors then present an experimental evaluation method (Section IIID) that can optionally be used to test if they should collect more data.
B. Notation
- Let the robot be R, the object be O, and the pushing strategy (or set of data) be D. Each sample of D contains the data for one robot push and the corresponding object movement: (px, py,rx,ry,sx,sy,θ) (Fig. 2).
- Θ represents the change in orientation of the object and is not drawn in the figure.
- The robot can be in contact with the object at multiple points.
- The basic strategy is to choose a random direction and distance (within certain limits) for the robot to push with from its current position.
- The collected data can be used immediately.
D. Experimental Evaluation of Existing Data
- After the authors collect some data with the above method, they can experimentally evaluate this data (Algorithm 1 Experimental Evaluation) to test whether or not it has enough samples for successfully solving goal queries (ie. pushing object to goal).
- If the authors can solve these goal queries with a good success rate, they have enough samples and can stop collecting data.
- In Algorithm 1 (Experimental Evaluation), ob ject not move is true if the robot pushes forward but the object does not move (ie. the robot has moved away from the object).
- The query fails if any of the conditions in lines 12 and 14 are true.
- The number of previous queries is a parameter.
E. Object Delivery by Using Collected Data
- The authors describe how to use the collected pushing strategy to control the robot to push the object towards the goal.
- The authors continuously perform this computation and execute the appropriate robot push until the object reaches the goal.
- The dist function computes the Euclidean distance between (px, py,sx,sy) and the corresponding values in each samplei.
- The runtime and storage space for Algorithm 2 are both in the order of the size of the pushing strategy.
- The control program continuously receives the markers’ positions/orientations.
V. EMPIRICAL EVALUATION
- Fig. 6 shows the robot and irregular-shaped objects that the authors used.
- An advantage of the initial method is that the authors can easily collect a variety of robot pushes with different positions and directions.
- The authors method has a better success ratio than the dipole method in all cases.
- For the L-shape, their method fails for two trials because the object went outside of the boundary.
Did you find this useful? Give us your feedback
Citations
253 citations
199 citations
145 citations
86 citations
Cites background from "Automatic learning of pushing strat..."
...Recently, there has been interest in generating push trajectories using sampling based planners (Cosgun et al., 2011; Lau et al., 2011), trajectory optimization (King et al....
[...]
...Recently, there has been interest in generating push trajectories using sampling based planners (Cosgun et al., 2011; Lau et al., 2011), trajectory optimization (King et al., 2013), and learning methods (Zito et al., 2012)....
[...]
80 citations
Cites background from "Automatic learning of pushing strat..."
...Recently, there has been interest in generating push trajectories using sampling based planners [18, 19] and learning methods [20]....
[...]
References
6,340 citations
704 citations
"Automatic learning of pushing strat..." refers methods in this paper
...For the second problem, we use a non-parametric regression technique to decide where to move the robot and which direction it should push in, given the final goal and the robot/object relative locations....
[...]
513 citations
"Automatic learning of pushing strat..." refers background in this paper
...In addition, the data implicitly stores relevant physical parameters such as the weight distribution and the friction between the robot, object, and pushing surface....
[...]
144 citations
"Automatic learning of pushing strat..." refers methods in this paper
...We then use this pushing strategy to push the object towards a user-specified goal....
[...]
135 citations
"Automatic learning of pushing strat..." refers methods in this paper
...For the second problem, we use a non-parametric regression technique to decide where to move the robot and which direction it should push in, given the final goal and the robot/object relative locations....
[...]
Related Papers (5)
Frequently Asked Questions (12)
Q2. What are the future works mentioned in the paper "Automatic learning of pushing strategy for delivery of irregular-shaped objects" ?
The authors may further use existing strategies and generalize them for different shaped objects in the future. In the future, the authors can take into account such cases with their pushing approach. For future work, the authors can define a formal measure of the variety of samples in a pushing strategy based on the robot positions and push directions. In addition, some pushing strategies may contain more data than the authors need, and they can explore the idea of pruning the samples that do not improve the result significantly.
Q3. what is the objective of the initial data collection process?
The objective of the initial data collection process is to start with no data, and collect a variety of robot pushes and object movement as quickly as possible.
Q4. What is the function that is used to send commands to the robot?
The host computer continuously tracks the position/orientation of the robot, object and goal with the camera, and wirelessly sends control signals to the robot.
Q5. What are the advantages of the experimental evaluation method?
The advantages of the experimental evaluation method are that the authors can test if the authors have enough data with the current pushing strategy, and the authors can collect supplemental data at the same time.
Q6. what is the purpose of the algorithm?
ForAlgorithm 1: Data Collection Initial Data Collection: for k = 1 to K do1 if dist(prev O pos,O.pos())< ε then2 reset robot();3 if ob j outside boundary() then4 reset robot ob j boundary();5 prev O pos = O.pos();6 R.spin(random angle());7 R. f orward(random dist());8 D.save sample();9Experimental Evaluation of Existing Data: reset robot();10 while true do11 if ob ject not move() or over time limit() then12 reset robot();13 if ob j outside boundary() then14 reset robot ob j boundary();15 solve query(random goal());16 if good success rate() then17 break;18the purpose of collecting data samples, the robot can be positioned anywhere as long as it is in contact with the object.
Q7. What is the advantage of the initial method?
An advantage of the initial method is that the authors can easily collect a variety of robot pushes with different positions and directions.
Q8. What is the purpose of the push plan?
Push plans for circular objects are computed that allow the object to touch and move along the wall/obstacles [2], [3], [4], or allow for multiple pushes of the object [4].
Q9. What is the simplest way to push a robot?
If there are obstacles, the authors first compute a global collison-free path with existing planning techniques [15], and then execute the robot to push the object along sub-goals of this path.
Q10. What is the way to test the push strategy?
The authors empirically show that their pushing strategy collection and robot control algorithms work well and, most importantly, are independent of the shape and weight distribution of the object.
Q11. How do they measure the object’s shape?
they explicitly measure the object’s shape by using a proximity sensor on a robot finger to detect a point cloud of the object, and then fit a shape to these points.
Q12. What are the inputs for the pushing strategy?
The inputs are the boundary of the workspace where the robot and object can move in, the radius of the circular robot, and the radius of the bounding circle of the object.