The concept of “Interactive Fleet Learning” (IFL) addresses a significant development in the field of robotics and artificial intelligence. Large groups of robots, or fleets, have emerged from laboratories to perform practical tasks in real-world settings. Examples of these include Waymo’s fleet of over 700 self-driving cars operating in several cities and the industrial application of robot fleets for tasks such as package delivery and e-commerce order fulfillment at companies like Amazon and Ambi Robotics.
These robots can operate autonomously in unstructured environments by using advanced deep learning technologies. They share their collective data, enabling the fleet to learn efficiently from individual experiences. This sharing and learning process is facilitated by storing data, memory, and computational details in the cloud via the internet, an approach known as “Fleet Learning”.
However, Fleet Learning encounters a challenge known as the “long tail”, where robots face new scenarios and edge cases that are not represented in the existing dataset. This uncertainty implies that future conditions cannot always mirror past situations, raising questions about the reliability of robotic services.
One solution is to rely on remote human operators who can step in and take control of the system when necessary, a process known as teleoperation. This method ensures the continual improvement of robot policies and reduces the need for human intervention over time.
While industry has embraced this model, it has not been widely adopted in academia, leading to a reliance on ad-hoc solutions to determine when robots should yield control. To address this, a paradigm known as Interactive Imitation Learning (IIL) has been introduced, where the robot intermittently gives up control to a human supervisor and learns from these interventions. However, determining when and how to switch between robot and human control remains an unresolved issue, especially when extended to the scale of robot fleets.
In response to this, a recent paper introduced Interactive Fleet Learning (IFL) as a paradigm for interactive learning involving multiple robots and multiple humans. It consists of four main components: on-demand supervision, fleet supervision, continual learning, and internet connectivity. The paper also explored the allocation of human supervision to large fleets and proposed the Fleet-DAgger, a family of IFL algorithms.
Finally, the researchers introduced the IFL Benchmark, an open-source Python toolkit for facilitating the development and evaluation of IFL algorithms. The aim is that these new formalisms, algorithms, and benchmarks will bridge the gap between the theory and practice of robot fleet learning and stimulate further research. The IFL paradigm may indeed pave the way towards scaling up supervised robot learning and deploying reliable robot fleets in the real world.
