170 likes | 388 Views
Multi-Modal Motion Planning for a Humanoid Robot Manipulation Task. Ye Fan. Overview. This paper presents a motion planner that enables a humanoid robot to push an object on a flat surface. Difficulties: 1. Current techniques can only plan in single modes.
E N D
Multi-Modal Motion Planning for a Humanoid Robot Manipulation Task Ye Fan
Overview This paper presents a motion planner that enables a humanoid robot to push an object on a flat surface. Difficulties: 1. Current techniques can only plan in single modes. 2. Traditional search-based method are inefficient. Solution: Random-MMP: randomly samples mode transitions to distribute a sparse number of modes across configuration space Test: Honda ASIMO robot
Why is push hard • Need to switch hands • Need to walk to new location • Need to Avoid collision • Configuration cannot be recovered • Different mode: WALK, REACH and PUSH • Multi-modal planning • Continuous set of neighbors Conclusion: Search samples modes too densely
Configurations Let C denote the configuration space. • qrobot has three dimension (xrobot, yrobot, θrobot). • qobject has three dimension (xobject, yobject,θobject). • Each arm has 5 joint angles. • Each hand has open and close status. So C is a 18 dimension spaces: 3 + 3 + 5 * 2 + 1 * 2 = 18
Modes and Submanifolds Each mode constrains motion to a sub-manifold of lower dimension: • Walking mode only change qrobot • Reach mode will only affect one arm and its hand We define that different modes are adjacent by: q belongs to Fm∧Fm’ q is a transition configuration A constraints: object must be in the view of robot’s eyes.
Configuration Space Paths from q to q’ pass through a transition configuration in Fm∧Fm’
Single-mode planning Problem: The set of all such configuration has zero measure in the 6D reach submanifold, so a randomly sampled arm configuration has zero probability of transitioning to a push.
Existing Multi-modal approaches We need mode-before-motion approaches. Details: • Pick up an unexpanded node from T. • For each adjacent mode m’, plan a single-mode path y in m, starting at a transition a’ in Fm∧Fm’. If successful, add the edge (q, m) -> (q’, m’) to T. Once goal is reached, the motion follows the single-mode motions along the edges of the solution path.
The Drawback of Search • Search is applicable when each mode has a finite number of adjacencies. • Search costs too much, since we need to discretize walk positions, contacts and pushes. 2 (hands) * 4 (sides of box) * 3 (motion of push, CW/CCW rotate) * 4 (walking positions) = 96 (branching factor) Conclusion: Search covers the configuration space much more densely.
Random-MMP Inspired from PRM which uses randomness to overcome similar discretization issues Original Random-MMP: Random-MMP maintains a tree T and extends it with a single-mode transition. Each extension picks a node from T at random with probability Φ, and expands to a single adjacent mode sampled at random with probability Ψ.
Random-MMP How to select Φ and Ψ? RRT-like implementation of Random-MMP: • Sample a random configuration qrand. • Pick a node (q, m) that minimizes a distance metric d(q, qrand). We define d to be the distance of the object configurations. (Φ indicated) • The tree is expanded from (q, m) to new mode(s) using an expansion strategy Expand.
Expansion Strategies • Blind: Expand to an adjacent mode m’ chosen at random. • Reach/Utility-Informed: Same, but samples contacts, for reach-to-push transitions, according to expected reachability/utiliy. • Push-centered: Expands a sequence of modes that executes a high utility push that move qobj toward qrand.
Reach/Utility-Informed Sampling A 3D workspace W of points (x, y, θ), where (x, y) are the horizontal coordinates of p and θ is the orientation of n. • Reachable checks whether hand can be moved to touch the object with desired orientation without any collision. • Utility estimates the distance that contact can be pushed in the absence of obstacle.
Push-centered expansion Maximum push utility depends greatly on the placement of the robot body. • Choose a robot’s body and arm configuration and a reach-to-push transition qpush. • Plan a whole sequence of modes backwards from the reach mode at qpush to (q, m). • Plan a path forward from qpush.
Simulation and Experiments Replanning in a changing environment
Future Work • Modify cameras to provide continuous feedback. • Planning and executing unstable pushes. • Improve planning speed. • Advance the understanding of muti-modal problems.