Ready, bid, go! On-demand delivery using fleets of drones with unknown, heterogeneous energy storage constraints
Highlights
- Overall, this was an easily approachable and enjoyable paper to read. The ideas are novel, and the authors take a unique approach to dealing with a decentralized drone system. However, I think there could have been just a bit more polish. I would have given this paper a weak accept.
Summary
UAV based logistical distribution is expected to help reduce delivery time, costs, and emissions. This paper deals with on-demand delivery scenarios, that is, where orders arrive stochastically and a drone needs to be assigned from a fleet to carry out that delivery. The authors introduce a decentralized deployment strategy that combines auction-based task allocation with online learning to allow each drone to "bid" on a delivery based on their perceived energy storage levels, parcel mass, and delivery distance. They show that counter-intuitively, adopting a winning bid policy of "least-confidence" reduces delivery times and increases the number of successful orders over time.
Key Contributions
Per the paper:
-
Decentralized learning-based deployment strategy for UAVs with unknown battery health
-
Long-term evaluation of heterogeneous fleets
-
Forecasting-enabled order commitment
Strengths
-
This paper is very approachable, Figure 2 was excellent for getting the holistic idea of their entire system early, and is easy to draw back to as one moves through the sections.
-
Simulation is an appropiate size and scale.
-
They address heterogenous swarms with unknown energy storage capacities, which is well-grounded in reality.
-
Their counter-intuitive findings, that a least-confident winning bid policy is the best bidding policy, is actually a very interesting and novel result.
Weaknesses / Questions
-
Moderate: From personal experience in this field, wind factor is can be a substantial factor for battery life (i.e. upwind traversal consumes substantially more energy than downwind). I don't believe that was mentioned until the very end, so I was carrying that doubt with me throughout the entire read. I believe they should have addressed that earlier.
-
Minor: What is the performance against other learning techniques? Or different baselines in general? It seems like it's really only evaluated against threshold in one figure (Figure 5).
-
Minor: It is done completely in simulation versus live deployement. The authors are forthcoming about this. Not a major weakness since their target hardware is prohibatively expensive, and I don't think that should degrade their interesting ideas and findings. It would however demonstrate just how much wind affects the drones.
-
Minor: Why doesn't MILP work here? Maybe an extra sentence or two would further boost their paper (e.g. "MILP is usually only used for centralized planning").
-
Nit: The learning techniques themselves are very simple but well established, which I don't believe is a weakness given the authors were creative in how they were applied. In fact, I believe their use makes their system more robust versus adopting more complicated techniques that just sound impressive, but run the risk of overfitting. What I believe the authors could have talked up more was how their simple techniques added to the overall explainability of the system. That would help boost trust in the system if it was ever deployed.
Related Work
-
Decentralized and Online Learning
-
MILP, Queuing theory
-
Auctions
-
Supervisory control theory