In a recent paper, Facebook and University of California, Berkeley researchers propose an approach to payload-carrying drone flight planning that “learns how to learn” models of environmental dynamics. Results from experiments suggest the work could inform the development of future robots — perhaps in warehouses or other industrial settings — that physically interact with and adapt to the world in the face of unpredictability.
The team used meta-learning, a subfield of machine learning where learning algorithms are applied on metadata about machine learning experiments, to train a model for fast adaptation to changing dynamics in the context of a suspended payload control task. In this task, a quadcopter had to position itself to pick up a target object along a path to a goal destination.
One of the biggest challenges stemmed from the variability different payloads introduced, all of which were attached via cable and a magnetic gripper. For example, a payload with a shorter cable oscillated faster compared with one attached with a longer cable.
To address this, the team trained a dynamics model with data from a range of physical conditions, like different payload masses and tether lengths, and augmented it with variables representing unknown environmental and task factors. This enabled the system to adapt to new payloads at test time by initializing the dynamics model, getting the current state, solving for an action, executing that action, recording the outcome, and then retraining the dynamics model.
The researchers collected the initial training data by having a person pilot the quadcopter (a DJI Tello) along random paths for each different suspended payload. (The payloads in question were 3D-printed boxes weighing between 10 and 15 grams.) Data including the controls and location of the payload, tracked with an externally mounted RGB camera, was recorded every 0.25 seconds and saved into a data set consisting of separate data sets — one per payload task.
The final corpus consisted of approximately 16,000 data points from 1.1 hours of flight, 5% of which was reserved for evaluation.
In the course of the experiments, the researchers reported, the quadcopter delivered payloads to their destinations the majority of the time. That said, they acknowledged there’s room for improvement; the time the suspended payload was picked up or dropped off had to be manually specified, and the approach only assumed an estimate of the suspended payload’s position. They left overcoming those challenges to future work.
“We believe this is the first meta-learning approach demonstrated on a real-world quadcopter using only real-world training data that successfully shows improvement in closed-loop performance compared to non-adaptive methods for suspended payload transportation,” wrote the researchers. “Although we only consider the specific task of quadcopter payload transportation in this work, we note that our method is general and is applicable to any robotic system that interact with the environment under changing conditions.”