Melvin's digital garden

Planning and control in belief space

Title: Planning and control in belief space

Types of uncertainty

  • error in measurement (small-scale)
  • no information on certain locations (large-scale)

Robot architecture: observation - state esimtation - action selection - action

State estimation via Bayesian belief update, where belief is a probability distribution over the states of the world.

Representation of belief: gaussian, histogram, set of particles, bayesian network

Optimistic (re)planning in belief space

Assumptions:

  • that are no bad outcomes
  • actions result in most likely transition and observations
  • replan if expectation is violated at runtime

Control with state-dependent observation noise

Belief space dynamics: specify next belief space as a function of previous belief state and action

State update: generalized Kalman filter

Substitute expected observation in for the actual one, add Gaussian noise.

Planning by local optimization.

Robot grasping with tactile sensing

Robot space: 11 DOF (fully observable)

Object pose: 3 DOF (partially observable)

Belief space is probability distribution over the DOF of the object.

Small number of macro actions, each action can be quite complicated.

Observations: arm trajectory according to proprioception, force sensors on the finger tips

Forward search to do planning. Prune and cluster.

Household robot with local observation

Uncertainty in the large.

Initial state is geometric detail of the house. Goal set is abstract, symbolic.

Hierarchical planning can help to keep the problem size small.

Links to this note