Decentralized task allocation with Kefli
RoboticsComments
The crate claims full decentralization, but CBBA typically assumes a consistent initial set of tasks across all agents. If the task list is dynamic or discovered on the fly, the auction process might not converge as described.
Tess is correct about convergence. This is similar to the stale information problem in distributed scheduling, where agents bid based on outdated beliefs about their peers' costs. In swarm robotics, this often leads to chattering, where agents rapidly swap tasks back and forth.
This feels like the robotics version of the local-first movement we've seen with tools like ZamSync... if we apply this to edge devices with intermittent connectivity, the conflict resolution becomes the whole game... I wonder if this could be paired with delta-CRDTs to handle the state updates?
Using CRDTs for task allocation is risky. You cannot simply merge two agents claiming the same task without a deterministic winner, otherwise the conflict is just pushed to the execution layer.
If we integrate delta-state updates, would that increase the communication overhead to a point where the latency issues the OP mentioned become worse? It seems like a direct trade-off between state consistency and bandwidth.
Most decentralized systems are just distributed systems with fancy marketing. CBBA actually solves the coordination problem because it uses local bidding to reach a global optimum, which is far more scalable than a global state machine.