← Back ICRA 2026

Distributed AI for Robotics

Satyabhama Singh, Lars Wulfert, Hendrik Woehrle

PDF

AI summary

Federated learning enables collaborative robot training but suffers performance degradation under non-IID data, with SAC+HER proving more stable than DDPG+HER.

federated learning deep reinforcement learning non-IID data robot manipulation multi-robot systems distributed training

Problem

Centralized robot training limits scalability and collaboration while overloading servers, whereas federated learning offers a distributed alternative but degrades under the non-IID data inherent to multi-robot environments.

Approach

The authors simulate a multi-robot manipulation task to compare centralized and federated training using SAC+HER and DDPG+HER, specifically measuring how heterogeneous goal distributions affect learning stability and success rates.

Key results

SAC+HER demonstrates stable training and higher success rates than DDPG+HER in both centralized and federated settings
DDPG+HER shows high sensitivity to exploration noise and lower overall success rates
Non-IID goal distributions degrade performance for both algorithms, causing increased instability in DDPG+HER
Varying exploration noise across federated clients improves DDPG+HER performance by enabling diverse exploration

Why it matters

Provides foundational insights for deploying scalable, collaborative robot learning systems in real-world environments where data heterogeneity is unavoidable.

Abstract

Robot learning primarily relies on centralized train- ing. While it provides the infrastructure, centralization limits par- allel and collaborative learning among robots and place significant computational load on the central server, indicating the need for federated learning (FL) in context of multi-robot training. How- ever, robots trained in a federated setup are subjected to non-in- dependent and identically distributed data (non-IID), resulting in degraded model performance. This extended abstract presents the current state of research aimed at improving robot learning under non-IID conditions in FL. In this regard, this work provides an initial comparative analysis of robot learning methods in central- ized and federated training setups, with an emphasis on the impact of non-IID data on learning behaviour in a simulation environ- ment. The results highlight the differences in learning stability across algorithms and present the influence of non-IID goal distri- butions on performance.

Index terms

Reinforcement Learning Machine Learning for Robot Control AI-Based Methods