Real-time perception-guided motion planning for a personal robot
read more
Citations
Semantic 3D Object Maps for Everyday Manipulation in Human Living Environments
STOMP: Stochastic trajectory optimization for motion planning
An open-source research kit for the da Vinci® Surgical System
Integrated task and motion planning in belief space
Towards autonomous robotic butlers: Lessons learned with the PR2
References
Planning Algorithms: Introductory Material
Probabilistic roadmaps for path planning in high-dimensional configuration spaces
RRT-connect: An efficient approach to single-query path planning
Randomized kinodynamic planning
Related Papers (5)
Frequently Asked Questions (21)
Q2. What have the authors stated for future works in "Real-time perception-guided motion planning for a personal robot" ?
To validate its efficiency, the authors showed the integration of 3D collision maps and semantic annotations created in real-time from sensed laser data with a motion replanning framework as building blocks towards efficient and robust pick and place scenarios. While this is still work in progress, as the authors plan to deploy and integrate their architecture with grasping planners to close the sensing-action loop, they already make their software available as an off-the-shelf component of the ROS project.
Q3. What is the purpose of the experiments in this paper?
For the purpose of the experiments in this paper, the authors are mostly interested in pick and place operations, and thus the most important areas that the authors need to detect are tables and other horizontal planes which can support objects.
Q4. What is the purpose of the 3D perception system?
Its purpose is to also provide contextspecific control, responsiveness to sensory events, and failure detection, analysis, and recovery.
Q5. What is the second actor of the 3D perception pipeline?
The second actor of the 3D perception pipeline is the Scene Interpreter node, which continuously segments and classifies surfaces by fitting planar models to them, in order to abstract sensor data and make it more informative.
Q6. What is the purpose of the experiment?
For the purposes of their experiments, constraints that restrict the position of certain robot parts (the gripper) at the goal state were all that was necessary.
Q7. What is the function that triggers the record static map service call?
The record static map service call triggers a special mode in the node which records a complete sweep map as a static fixed reference map, that all recorded data will be merged with until the next service call.
Q8. What is the way to make the robot operate at high speeds?
If the robot is to be allowed to operate and manipulate at high speeds, the laser tilt unit must be actuated as fast as possible, to account for the environment dynamics.
Q9. What is the version of KPIECE that is presented in Algorithm 3?
Currently the library contains implementations of RRT (Rapidly-exploring Random Trees) [14], [15], EST (Expansive Space Trees) [16], [17], SBL (Single-query Bi-directional probabilistic roadmap planner with Lazy collision checking) [18] and the version of KPIECE presented in Algorithm 3.
Q10. What is the way to create a static obstacle map?
Having a static reference map is beneficial during longer manipulation scenarios, where the estimated collision map may contain holes due to occlusions caused by the robot’s arms at certain positions in time.
Q11. What is the purpose of the PR2?
A system as complex as the PR2 is driven by a number of subsystems that need to easily be able to communicate with each other: hardware drivers, controllers, visual perception, motion planning, higher level control, etc.
Q12. What is the main idea behind motion planning?
In order to address these needs, sampling-based motion planners, which have been shown to be able to handle such problems [8], [9], are used.
Q13. What is the main idea behind sampling-based motion planning?
The main idea behind sampling-based motion planning [8], [9] is to sample the state space X of the robotic system and maintain a data-structure of the samples that can be used to guide the robot to the goal [13].
Q14. What is the definition of a solution to a motion planning problem?
A solution to a motion planning problem is defined to be the set S = {x0, ...,xk}, where x0 is the initial state and xk is in the goal region.
Q15. What is the definition of a constraint?
In general, a constraint is a function f : X → {true, false}, meaning that for any state x ∈ X , it can be decided whether it satisfies a constraint or not.
Q16. What is the way to calculate the axis of the robot base?
When moving the robot base, without moving the arms, the authors employ a grid-based global planner and a local controller designed for efficient motion in tight spaces.
Q17. Why is the object in the gripper incomplete in P?
Please note that the object present in the gripper from the left part of Figure 4 is incomplete in P , because parts of it are removed by the robot self filter at the laser scan level, due to the convex hulls of the gripper being much larger than the actual gripper itself.
Q18. What is the deciding factor on the average computation time per map?
A deciding factor on the average computation time per map is represented by the number of data leaves that are requested from the motion planner, that is, the resolution of the world.
Q19. What is the purpose of the Dynamic Obstacle Map?
Because there is no good fixed resolution that best accounts for all possible scenarios, the Dynamic Obstacle Map architecture is designed to be parameterizable on the fly.
Q20. What is the OMPL code that can do inverse kinematics?
In addition to planning algorithms, OMPL contains code that can do inverse kinematics using a genetic algorithm: GAIK (Genetic Algorithm Inverse Kinematics) [19].
Q21. How long does the system take to complete its computations?
As shown, both for laser sweeps with a period of 1 or 2 seconds, all components manage to finish their computations in real-time, that is before the next state update.