Robot team learning method to achieve 8.8 times human return

When robots are at risk or unable to complete tasks , Commercial and industrial deployments of robot fleets often rely on remote human remote operators during execution . Through continuous learning , Interventions from remote crowds can also be used to improve robot fleet control strategies over time . A core problem is how to effectively allocate limited human attention to a single robot . Previous work in single robot 、 This problem is solved in a single person environment . We will learn from the interactive team (IFL) Set formalization , Among them, multiple robots interactively query and learn from multiple human supervisors . We provide a fully implemented open source IFL Benchmark Suite , This kit contains GPU Accelerated Isaac Gym Environmental Science , Used to assess IFL Algorithm . We proposed Fleet-DAgger, One IFL Family of algorithms , And a novel Fleet-DAgger Algorithm and simulation 4 Two baselines were compared . We also use 4 individual ABB YuMi The robot arm carried out 1000 Physics push block experiment . Experiments show that , Assigning humans to robots will significantly affect the performance of robot fleets , And our algorithm can achieve higher than the baseline 8.8 Times the human return .