AI Balances Speed and Sustainability in Robot Recovery

In industrial settings where robots work alongside humans, unexpected disruptions like lighting failures or cyberattacks can cripple performance, forcing systems to rely heavily on human intervention and consume more energy. A new study introduces an AI-driven approach that helps these collaborative robots recover more quickly and efficiently, balancing the need for rapid response with environmental sustainability. The research, conducted using a real-world collaborative robot named CORAL, demonstrates how intelligent decision-making can reduce recovery time and energy use, though it also uncovers a concerning tendency for robots to "forget" their original training after disruptions.

The core of this research is the GResilience framework, which employs three types of AI agents to guide robots during recovery from disruptive events. These agents include one that uses optimization to weigh trade-offs, another that applies game theory to model competing goals, and a third that learns through reinforcement to adapt over time. In experiments, the reinforcement learning agent proved most effective, enabling the robot to recover up to 17% faster than its internal policies alone, while also increasing autonomous actions by 27%. However, this improvement came at a cost: the AI agents required additional computational power, leading to higher carbon emissions during the recovery process.

To evaluate these policies, the researchers developed a resilience model that tracks system performance through a metric called the Autonomous Classification Ratio (ACR), which measures how often the robot acts independently versus needing human help. By analyzing ACR over time, the model identifies three key states: steady operation, disruption, and recovery. This allows for automatic detection of performance drops and precise timing of interventions. The framework was tested in over 800 simulated scenarios and four real-world experiments, where disruptions such as darkened lighting or manipulated image data were introduced to degrade the robot's object-classification abilities.

Showed that all AI-based policies outperformed the robot's internal decision-making, with the reinforcement learning agent achieving the best balance between speed and autonomy. For instance, in simulations of a lighting failure, the reinforcement learning agent reduced the recovery state duration by 14% compared to the internal policy, while increasing autonomous actions from 64% to 81%. However, the study also revealed a significant limitation: after recovering from a disruption, the robot often experienced "catastrophic forgetting," losing memory of its original training and requiring additional human input to relearn tasks, which could undermine long-term efficiency.

Beyond decision-making, the research explored containerization as a to enhance both resilience and greenness. By packaging the robot's software components into lightweight containers, the system reduced energy consumption by 50% compared to a traditional multi-machine setup, cutting CO2 emissions from 0.198 to 0.099 kilograms of CO2 equivalent over two hours of operation. This approach not only saved energy but also improved system resilience by allowing easier replication and management of components, though it introduces complexity in deployment and maintenance.

Despite its successes, the study has limitations. The reliance on a single case study, CORAL, may restrict generalizability to other collaborative systems, and the increased CO2 emissions from AI computations highlight a trade-off between intelligence and sustainability. Future work could focus on refining the AI agents to minimize energy overhead, exploring more diverse disruption scenarios, and integrating the decision-making tools into broader cyber-physical system architectures to enhance real-world applicability.