Decentralized Sharing and Valuation of Fleet Robotic Data

Yuchong Geng¹ Dongyue Zhang¹ Po-han Li² Oguzhan Akcin² Ao Tang¹ Sandeep P. Chinchali²

¹Cornell University ²The University of Texas at Austin

TL;DR: Robots deployed across different environments observe very different data. We propose a decentralized, privacy-aware framework in which robots learn to predict which local experiences are valuable to other robots in the fleet, selectively share only a small fraction of that data, and improve learning through peer feedback.

System diagram for decentralized data sharing across a robotic fleet

Decentralized sharing: robots predict which local experiences are valuable to peers, share selectively under privacy constraints, and use feedback to improve.

Overview

Robotic fleets deployed across cities, homes, and roads experience highly heterogeneous environments. Conditions that are rare and safety-critical for one robot (e.g., snow, unusual traffic patterns, or construction zones) may be commonplace for another. Aggregating all sensory data centrally is costly, slow, and often infeasible due to bandwidth, privacy, and organizational constraints.

We propose a decentralized, peer-to-peer framework that enables robots to proactively share only the most valuable local experiences with their peers. Each robot learns a global utility (GU) sharing model that predicts how useful each observation will be to other robots, rather than relying on centralized coordination. A local privacy filter limits sensitive content, and peers provide feedback (e.g., out-of-distribution signals) that improves valuation and sharing over time.

Core idea

Selective sharing: robots transmit only a small, high-utility fraction of their data.
Predictive data valuation: a learned GU model estimates how valuable each observation will be to peers.
Decentralized learning: sharing decisions are made locally without a central coordinator.
Privacy by design: a local filter removes or obfuscates sensitive information before sharing.
Feedback-driven improvement: peer OoD signals refine future sharing decisions.

How it works

In each round, a robot evaluates local observations using its GU model to predict their utility to peers, shares only the top-ranked subset after privacy filtering, and receives feedback that improves future valuation.

Local inference: a robot processes a new observation with its task model and produces embeddings and predictions.
Utility estimation: the GU model predicts how valuable the observation will be to peers.
Selective sharing: only top-ranked data is transmitted after privacy filtering.
Peer feedback: recipients score shared data via OoD detection and return feedback.
Adaptive valuation: feedback refines the GU model over successive rounds.

Figures

Toy experiment accuracy over sharing rounds

Toy experiment: learned data valuation rapidly improves model accuracy across sharing rounds, outperforming random sharing and approaching an oracle upper bound.

Why this matters today

Modern robotic fleets, including autonomous vehicles, are deployed across diverse cities and environments. Rare but safety-critical conditions such as snow, unusual road layouts, and long-tail edge cases are unevenly distributed across the fleet.

This work anticipates the need for decentralized mechanisms that allow robots to discover, value, and exchange such experiences efficiently without centralized data aggregation. As large-scale robot deployments accelerate, selective and privacy-aware data sharing becomes a key bottleneck for continual learning at scale.

Open research questions

Data valuation: how should we fairly price data contributions in a decentralized fleet?

Privacy-utility trade-offs: how can robots share useful data while protecting sensitive information?

Convergence: when does the sharing protocol reach a steady state across peers?

Discovery at scale: can robots quickly find peers likely to have valuable data?

BibTeX

@InProceedings{pmlr-v164-geng22a,
  title = 	 {Decentralized Sharing and Valuation of Fleet Robotic Data},
  author =       {Geng, Yuchong and Zhang, Dongyue and Li, Po-han and Akcin, Oguzhan and Tang, Ao and Chinchali, Sandeep P.},
  booktitle = 	 {Proceedings of the 5th Conference on Robot Learning},
  pages = 	 {1795--1800},
  year = 	 {2022},
  editor = 	 {Faust, Aleksandra and Hsu, David and Neumann, Gerhard},
  volume = 	 {164},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {08--11 Nov},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v164/geng22a/geng22a.pdf},
  url = 	 {https://proceedings.mlr.press/v164/geng22a.html}
}

Contact

Questions? Email yg534@cornell.edu or sandeepc@utexas.edu.