An open-ended, goal conditioned robotic manipulation benchmark tackling industry's problems.
Jack Collins*, Mark Robson*, Jun Yamada*, Karol Janik, Mohan Sridharan, Ingmar Posner
Available robotic manipulation benchmarks all suffer from limitations, primarily their lack of industry engagement and inability to extend to more challenging environments. We present a new benchmark that was designed with challenges faced by industry in mind along with open-ended goal conditioning that allows the benchmark to cover a broad spectrum of challenging situtations.
Informed by industry, packaged for researchers
Diverse goal configurations
Accessible - Cheap and easy to create
Detailed simulated environment
Complex physical interactions
Task Planning
Hierarchical learning and planning
Challenging object state estimation
The parts to the benchmark are printed on a simple 3D FDM printer, cut from 20x20mm extruded profile aluminium, printed from an office printer or easily purchased making the parts cheap and easily accesible worldwide.
Check back soon for a link!
Baseline performance of traditional task planning and manipulation skills is presented, with code to be made available to be extended upon.
Check back soon for a link!
A high fidelity simulation environment of the benchmark has been developed in Nvidia Isaac Sim to facilitate both development, accessibility and learning-based approaches.
Check back soon for a link!
Jack Collins
Mark Robson
Jun Yamada
Karol Janik
Mohan Sridharan
Ingmar Posner
See the contact us page.