Integrated task and motion planning in belief space
read more
Citations
A Concise Introduction to Decentralized POMDPs
Integrated Task and Motion Planning
Incremental Task and Motion Planning: A Constraint-Based Approach
Planning in the continuous domain
Inner Monologue: Embodied Reasoning through Planning with Language Models
References
Probabilistic Robotics
Planning Algorithms: Introductory Material
Robot Motion Planning
Unscented filtering and nonlinear estimation
Related Papers (5)
Strips: A new approach to the application of theorem proving to problem solving
Frequently Asked Questions (13)
Q2. What future works have the authors mentioned in the paper "Integrated task and motion planning in belief space" ?
The additional operators to extend to the kitchen domain were relatively easy to construct. It is an important area of future work to explore more highly decomposed and non-analytic ( particlebased ) representations of belief, and to characterize operator pre-images in those representations. This is a critical area of future research, but the authors are optimistic that it will be possible to represent a wide variety of types of uncertainty over very large domains. This integrated general-purpose mechanism current supports robust, flexible, solution of simple mobile manipulation problems in the current implementation, and the authors expect it to serve as a foundation for the solution of significantly more complex problems in the future.
Q3. Why do the authors want to keep the basic structure of planning?
Because the authors are interested in regression-based planning using logical representations of sets of world states, the authors wish to retain the basic structure of planning to achieve a goal, rather than optimizing the sum of state and action costs over a fixed finite horizon or over the infinite horizon with value discounting.
Q4. What is the probability of a plan reaching a goal state?
In the probabilistic case, as long as each plan that is constructed has a finite probability of success, then eventually a goal state will be reached.
Q5. What is the approach to planning in domains with probabilistic dynamics?
The decision-theoretic optimal approach to planning in domains with probabilistic dynamics is to make a conditional plan, in the form of a tree or policy, supplying an action to take in response to any possible outcome of a preceding action.
Q6. What is the way to make planning with simplified domain models robust?
Planning with simplified domain models is efficient and can be made robust by detecting execution failures and replanning online.
Q7. What is the idea of characterizing a distribution in terms of pnm and computing?
The idea of characterizing a distribution in terms of pnm and computing its regression applies more generally, for example, to non-parametric distributions represented as sets of samples.
Q8. What is the approach to handling probabilistic uncertainty in the outcomes of actions?
Their approach to handling probabilistic uncertainty in the outcomes of actions is to: construct a deterministic approximation of the domain, plan a path from the current state to a state satisfying the goal condition; execute the first step of the plan; observe the resulting state; plan a new path to the goal; execute the first step; etc.
Q9. How many look operations are needed to rule out a move?
Continuing the search, the authors find that at least one look operation is necessary before attempting to move the object, in order to establish that the object is in the starting location of the move.
Q10. What is the way to achieve believing the object is in location 0?
There is only one way to achieve believing the object is in location 0 with high confidence, which is to look in location 0 to verify that the object is there.
Q11. What does the interleaved hierarchical planning and execution architecture do?
Interleaved hierarchical planning and execution fits beautifully with information gain: the system makes a high-level plan to gather information and then uses it, and the interleaved hierarchical planning and execution architecture ensures that planning that depends on the information naturally takes place after the information has been gathered; and•
Q12. What is the definition of pre-images in a continuous space?
In a continuous space, pre-images may be characterized geometrically: if the goal is a circle of locations in x, y space, then the operation of moving one meter in x will have a pre-image that is also a circle of locations, but with the x coordinate displaced by a meter.
Q13. What is the odometry error model used to represent the distribution over the product space?
This approach is convenient because the product of multiple tangent spaces together with regular real spaces for the other dimensions can be taken, and a single multivariate Gaussian used to represent the entire joint distribution over the product space [Fletcher et al., 2003].