Informed Hierarchical Physics-Based Manipulation Planning

We propose a hierarchical physics-based planning algorithm. The key idea is to use a computationally cheap coarse physics model to get a rough initial plan, then refine the plan sequentially with finer and more expensive physics models. A plan generated with a coarse model may not be feasible when checked with a fine physics model. However, it may yet contain part of a feasible solution.

 

We propose an informed hierarchical planner that identifies such a feasible part of the solution and defines a new planning problem at the next hierarchical planning stage.


We find through experiments in simulation and on a real robot that the informed hierarchical planner in comparison with a standard planner is one order of magnitude faster and is significantly more successful in finding feasible manipulation plans given a time limit. 

Combining Coarse and Fine Physics for Manipulation

Fast physics predictions can enable us realize real-time reactive object manipulation. 

We combine coarse (i.e. fast but inaccurate) physics models, with fine (i.e. slow but accurate) physics models, to generate hybrid models that are both fast and accurate enough for real-world manipulation. 

Our approach is based on parallel-in-time integration. 

 

Using our hybrid models for planning and control, we have shown significant speedups for real-world manipulation tasks without sacrificing success rates. 

Please see our ISRR 2019 paper. 

Task-adaptivity in Manipulation 

As humans, when we interact with objects sometimes we act fast, and other times we use slow actions in order to avoid undesired events.

How can a robot exhibit such an adaptive behaviour ? 

In this work, we model the problem as a Markov decision process (MDP) with action-dependent uncertainty in the state transition function. 

We consider a model where the uncertainty is proportional to the pushing speed. We provide an online solution to the MDP in real-time.

Moreover, our experiments show that a robot can exhibit an adaptive behavior during non-prehensile manipulation: pushing slow for high accuracy tasks and pushing fast for tasks that permit inaccuracy.

 

Please see our WAFR 2018 paper.

Online Re-planning for Manipulation in Clutter

Our goal is to grasp a target object in a cluttered environment as shown in the video.  

 

Previous work has focused on open-loop planning followed by blind execution (without feedback). This approach can easily lead to task failures due to uncertainty. A major reason for not re-planning continuously during execution is the long planning times this domain requires. 

To address this problem, we propose a stochastic trajectory optimizer and use it within a model predictive control (MPC) setting for manipulation in highly cluttered scenes. 

For more details see our IEEE Humanoids 2018 paper.