| <Poster Width="1735" Height="1227"> |
| <Panel left="-1" right="186" width="483" height="1035"> |
| <Text>Motivation:</Text> |
| <Text>Robots face two challenges in natural environments:</Text> |
| <Text>● Underspecified goals: no human to specify exact</Text> |
| <Text>goal configuration</Text> |
| <Text>● Uncertain dynamics: Effects of robot's actions on</Text> |
| <Text>novel objects is uncertain</Text> |
| <Text>Approach:</Text> |
| <Text>● For underspecified goals:</Text> |
| <Text>• Pose task as a constrained optimization problem</Text> |
| <Text>over a set of reward or cost terms.</Text> |
| <Text>• Can be defined manually or modeled from human</Text> |
| <Text>● For uncertain dynamics:</Text> |
| <Text>• Quickly approximate dynamics for a set of actions</Text> |
| <Text>• Plan efficiently using sampling-based techniques</Text> |
| <Text>Our algorithm:</Text> |
| <Text>● Searches in object configuration space using</Text> |
| <Text>Rapidly-exploring Random Trees (RRT)</Text> |
| <Text>● Adds leaves to search tree by forward-simulating</Text> |
| <Text>the learned dynamics for each object-action pair</Text> |
| <Text>● Uses directGD heuristic to quickly search</Text> |
| <Text>optimization landscape</Text> |
| <Text>● Returns a plan from the starting state to the</Text> |
| <Text>most optimal reachable state, given cost function</Text> |
| <Figure left="40" right="895" width="403" height="316" no="1" OriWidth="0.400578" OriHeight="0.244861 |
| " /> |
| </Panel> |
|
|
| <Panel left="483" right="187" width="827" height="589"> |
| <Text>Manipulation under uncertainty</Text> |
| <Text>Initial state</Text> |
| <Text>● Robot begins with a workspace containing </Text> |
| <Text>three unfamiliar objects</Text> |
| <Text>● Robot provided a cost function expressing </Text> |
| <Text>the following desiderata:</Text> |
| <Text>• Orthogonality</Text> |
| <Text>• Cicumscribed area</Text> |
| <Text>• Distance from edge of workspace</Text> |
| <Figure left="508" right="465" width="364" height="280" no="2" OriWidth="0.186127" OriHeight="0.181412 |
| " /> |
| <Text>Solution</Text> |
| <Text>● All objects pushed to orthonormal</Text> |
| <Text>orientations in the center of the workspace</Text> |
| <Text>● All paths free of collisions and redundant</Text> |
| <Text>actions</Text> |
| <Text>● Robot monitored error and replanned as</Text> |
| <Text>necessary</Text> |
| <Figure left="888" right="412" width="413" height="332" no="3" OriWidth="0.380347" OriHeight="0.183199 |
| " /> |
| </Panel> |
|
|
| <Panel left="486" right="778" width="828" height="441"> |
| <Text>Generalization to other manipulation tasks</Text> |
| <Text>Appropriate for tasks naturally expressed as</Text> |
| <Text>optimization of a cost function:</Text> |
| <Text>● Arranging clutter on a surface</Text> |
| <Text>● Multiple object placement</Text> |
| <Text>● Table setting</Text> |
| <Figure left="520" right="995" width="328" height="216" no="4" OriWidth="0.176301" OriHeight="0.100536 |
| " /> |
| <Figure left="873" right="866" width="428" height="343" no="5" OriWidth="0.383237" OriHeight="0.208222 |
| " /> |
| </Panel> |
|
|
| <Panel left="1318" right="188" width="407" height="676"> |
| <Text>Model Learning</Text> |
| <Text>Goal: discover the dynamics of each object</Text> |
| <Text>class over a set of action primitives</Text> |
| <Figure left="1342" right="296" width="366" height="191" no="6" OriWidth="0.30578" OriHeight="0.124218 |
| " /> |
| <Figure left="1335" right="500" width="380" height="195" no="7" OriWidth="0.744509" OriHeight="0.181859 |
| " /> |
| <Figure left="1366" right="709" width="315" height="143" no="8" OriWidth="0.406936" OriHeight="0.118409 |
| " /> |
| </Panel> |
|
|
| <Panel left="1318" right="873" width="406" height="346"> |
| <Text>Advantages</Text> |
| <Text>Appropriate for tasks naturally expressed as</Text> |
| <Text>optimization of a cost function:</Text> |
| <Text>• Unlike conventional single-shot methods,</Text> |
| <Text>doesn't require user specified goals</Text> |
| <Text>• Always guaranteed to return reachable solution</Text> |
| <Text>• Favorable anytime characteristics</Text> |
| <Text>• Feasible for real-time planning in high DOF problems</Text> |
| <Text>Similar to Reinforcement Learning formalism, but</Text> |
| <Text>trades path optimality for realtime feasibility</Text> |
| <Text>• RL can require many full-passes through</Text> |
| <Text>configuration space to converge to optimal policy</Text> |
| <Text>• Handling continuous features requires discretization,</Text> |
| <Text>tiling, or other appoaches</Text> |
| <Text>• RL better suited for problems with sparse reward</Text> |
| <Text>landscape, but optimizations offer a gradient (like</Text> |
| <Text>shaping reward) which allows fast heuristic search</Text> |
| <Text>with RRT</Text> |
| </Panel> |
|
|
| </Poster> |
|
|
