CoSTAR was developed as a system for quickly creating powerful, reusable robot programs with a broad range of capabilities. We use CoSTAR to collect data for the training of deep neural networks to complete complex tasks.
We developed CoSTAR with small manufacturing entities (SMEs) in mind. SMEs commonly have short turnover between products, and commonly need to retool production once or twice a week.The system combines our state-of-the-art perception system to detect objects and compute their poses with other capabilities for object tracking, motion planning, and reasoning. End-users can employ a Behavior Tree-based user interface to author task plans in a process that is as powerful as traditional programming.
The robot can use tools and collaborate alongside humans. Users can also provide it with high-level commands, like “pick up objects on the left side of the table.”
Check out CoSTAR:
- github.com/jhu-lcsr/costar_plan Deep Learning Planning Code
- CoSTAR Dataset Block Stacking for the evaluation of multiple input deep neural networks.
- github.com/cpaxton/costar_stack System Code for CoSTAR
- Faculty: Greg Hager, Marin Kobilarov, Darius Burshka
- Students: Andrew Hundt
People Previously Involved:
- Students: Chris Paxton, Felix Jonathan, Matt Sheckells, Chi Li
- Postdocs: Kelleher Guerin