Searching Behavior of a Simple Manipulator only with Sense of Touch Generated by Probabilistic Flow Control
Download as PPTX, PDF0 likes1,060 views
This document describes a study that uses probabilistic flow control (PFC) to generate search behavior in a simple robotic manipulator tasked with finding and grasping a fixed rod. PFC weights particles representing possible rod locations based on a value function, guiding the robot's motion in a search-like pattern as it taps and changes the direction of its arms while compensating for uncertainty in localizing the rod through touch sensing alone. The results demonstrate that PFC allows the robot to successfully complete the task through searching, while avoiding local minima issues, though longer completion times occur with higher weighting rates that prioritize exploration over exploitation. Future work will apply PFC to more practical cases and allow the weighting rate to vary.
![our past work1: real-time Q-MDP [Ueda 2005]
• method
• preparing a value function (a precise potential function)
beforehand under the assumption of no uncertainty
• calculating the expected improvement of the potential
with particles of MCL (Monte Carlo localization)
• decision making with consideration
of localization uncertainty
• certain -> keeping the position
in the goal
• uncertain -> rushing to the ball
Dec. 13, 2018 ROBIO2018 3
x 2
local minima problem unsuitable for simple navigation](https://image.slidesharecdn.com/20181213robio-181215033434/85/Searching-Behavior-of-a-Simple-Manipulator-only-with-Sense-of-Touch-Generated-by-Probabilistic-Flow-Control-3-320.jpg)
![our past work2:
probabilistic flow control (PFC) [Ueda 2015]
• giving large weights to particles that have good values
• reduction of local minima
• generating a kind of search behavior
Dec. 13, 2018 ROBIO2018 4
x
value
goal
deadlock
particles
real-time Q-MDP
x
goal
PFC
weighted by
the value](https://image.slidesharecdn.com/20181213robio-181215033434/85/Searching-Behavior-of-a-Simple-Manipulator-only-with-Sense-of-Touch-Generated-by-Probabilistic-Flow-Control-4-320.jpg)
![• Parameters of the previous works cause local minima.
• Large 𝑚 values prevent the local minima problem.
• Large 𝑚 values delay the task completion.
success rates
effect of the rate of weights 𝑚
Dec. 13, 2018 ROBIO2018 10
Q-MDP
PFC [Ueda 2015]
number of steps
large priority](https://image.slidesharecdn.com/20181213robio-181215033434/85/Searching-Behavior-of-a-Simple-Manipulator-only-with-Sense-of-Touch-Generated-by-Probabilistic-Flow-Control-10-320.jpg)
Searching Behavior of a Simple Manipulator only with Sense of Touch Generated by Probabilistic Flow Control
- 1. Searching Behavior of a Simple Manipulator only with Sense of Touch Generated by Probabilistic Flow Control Ryuichi Ueda Chiba Institute of Technology
- 2. decision making under uncertainty • The real world is essentially uncertain. • sensor noises • situations where some sensors are invalid • ex.: RGB camera in darkness • Animals and human beings decide their actions in uncertain situations. • ex.: a person who goes to his/her bedroom in darkness Dec. 13, 2018 ROBIO2018 2 problem: how to model such intelligent behavior
- 3. our past work1: real-time Q-MDP [Ueda 2005] • method • preparing a value function (a precise potential function) beforehand under the assumption of no uncertainty • calculating the expected improvement of the potential with particles of MCL (Monte Carlo localization) • decision making with consideration of localization uncertainty • certain -> keeping the position in the goal • uncertain -> rushing to the ball Dec. 13, 2018 ROBIO2018 3 x 2 local minima problem unsuitable for simple navigation
- 4. our past work2: probabilistic flow control (PFC) [Ueda 2015] • giving large weights to particles that have good values • reduction of local minima • generating a kind of search behavior Dec. 13, 2018 ROBIO2018 4 x value goal deadlock particles real-time Q-MDP x goal PFC weighted by the value
- 5. search behavior by PFC • motion that compensates for incomplete self-localization Dec. 13, 2018 ROBIO2018 5 only one landmark goal The robot may exist in this area. The robot must go here. (Is it possible??) robot: dragged by particles → search behavior
- 6. purpose •to find another search behavior • with another type robot • evaluation of various rates of weights Dec. 13, 2018 ROBIO2018 6 x goal low rate x goal high rate
- 7. searching rod problem • a simple robot manipulator and a fixed rod in the environment • task of the robot: to get the rod in its hand • restrictions of observation • When the robot touches the rod, it feels the rod somewhere in its body. • When the rod enters in the hand, the robot notices the task completion. • The angles of the joints are known. Dec. 13, 2018 ROBIO2018 7 x y fixed rod (unknown position)
- 8. applying PFC • off-line motion planning with known positions of the rod • solving V(𝜽, 𝒙rod) • V(𝜽, 𝒙rod): number of steps to the goal • 𝜽: joint angles (two dimensional, known) • 𝒙rod: position of the rod (two dimensional, treated as known) • on-line • localization of a rod with the touch sense • decision making • 𝑎 = argmin 𝑎 𝒙rod 𝑃(𝒙rod) V(𝜽, 𝒙rod) 𝑚 V 𝜽′, 𝒙rod Dec. 13, 2018 ROBIO2018 8 probability distribution where the rod is action of the robot posterior joint angles by 𝑎rate of weight
- 9. generated motion • known rod position • just the shortest time motion • PFC (unknown rod position): • The robot shows behavior like • searching the rod • tapping the rod • changing the folding direction of the arms Dec. 13, 2018 ROBIO2018 9 optimal control with the known rod position (background red color: probability distribution of the rod’s position)
- 10. • Parameters of the previous works cause local minima. • Large 𝑚 values prevent the local minima problem. • Large 𝑚 values delay the task completion. success rates effect of the rate of weights 𝑚 Dec. 13, 2018 ROBIO2018 10 Q-MDP PFC [Ueda 2015] number of steps large priority
- 11. conclusion • Searching behavior of a manipulator can be generated with PFC. • future works • applying PFC to more practical cases • making 𝑚 variable Dec. 13, 2018 ROBIO2018 11