Learning to Compose Hierarchical Object-Centric Controllers for Robotic Manipulation
Mohit Sharma*, Jacky Liang*, Jialiang (Alan) Zhao, Alex LaGrassa, Oliver Kroemer
* Equal Contribution[arXiv]
Embedded Files
Videos
Videos
Below we show learned policies for all tasks using all the different baselines and our proposed ExpandMDP based approaches.
Block Fit Task
Block Fit Task
Config: 1
Config: 1
End-Effector Action Space
One Controller
3-Priority
Proposed: 3-ExpandMDP-Single 3-ExpandMDP-Multi
Test Config 2
Test Config 2
End-Effector Action Space
One Controller
3-Priority
Proposed: 3-ExpandMDP-Single 3-ExpandMDP-Multi
Block Push Task
Block Push Task
Test Config: D
Test Config: D
End-Effector Action Space
One Controller
3-Priority
Proposed: 3-ExpandMDP-Single 3-ExpandMDP-Multi
Test Config: H
Test Config: H
End-Effector Action Space
One Controller
3-Priority
Proposed: 3-ExpandMDP-Single 3-ExpandMDP-Multi
Franka Hex Screw
Franka Hex Screw
End-Effector Action Space
One Controller
3-Priority
Combo 3-ExpandMDP-Single 3-ExpandMDP-Multi
Franka Door Open
Franka Door Open
End-Effector Action Space
One Controller
3-Priority
3-Combo 3-ExpandMDP-Single 3-ExpandMDP-Multi
Real World Experiments
Real World Experiments
Franka Hex-Screw
Franka Hex-Screw
One Controller
One Controller
3-Priority
3-Priority
3-Combo
3-Combo
3-Expand-Single
3-Expand-Single
3-Expand-Multi
3-Expand-Multi
Frank Door Open
Frank Door Open
One Controller
One Controller
3-Priority
3-Priority
3-Combo
3-Combo
3-Expand-Single
3-Expand-Single
3-Expand-Multi
3-Expand-Multi
Controller Selection Visualization Results: Block Fit
Controller Selection Visualization Results: Block Fit
The above figures plot the controllers selected by the 3-ExpandMDP-Multi method for the Block Fit task. As seen in the above figures the proposed approach uses object-axes controllers associated with different objects in parallel. More discussion for these results are in the Appendix of the paper.
Page updated
Google Sites
Report abuse