New AI method helps human operators regain control of vehicles after sudden disruptions

Imperial researchers have developed a fast, onboard learning system designed to help human operators quickly regain control of robots and vehicles when unexpected damage or environmental disturbances occur.

Robots and autonomous vehicles are increasingly deployed beyond tightly controlled settings such as warehouses and factories. However, when operating in the real world, sudden changes — including flat tyres, reduced motor power, uneven weight distribution or strong wind gusts — can cause vehicles to behave unpredictably. 

In extreme cases, this can overwhelm operators and lead to loss of control, with potentially serious consequences. In 2021, one of the world’s largest container ships, the Ever Given, became wedged in the Suez Canal in Egypt, after its operators lost control in high winds.  

In a new study published in Nature Communications, researchers from Imperial College London present Fast Learning‑Based Adaptation for Immediate Recovery (FLAIR), a learning-based system designed to support human operators when vehicles become difficult to control. 

FLAIR is designed to support operators in exactly these moments — when a vehicle is still functional, but no longer behaves as expected. By learning online, in milliseconds, how these disturbances affect the system, FLAIR helps restore intuitive control without requiring new controllers or lengthy retraining." Dr Maxime Allard

Rather than replacing existing control systems, FLAIR works as an assistance layer. It continuously learns how damage or environmental changes affect a vehicle’s behaviour and automatically adjusts the operator’s commands in real time, helping the vehicle respond as expected again. Crucially, the system operates entirely onboard, without the need for external computing or advance retraining. 

FLAIR updates its internal model every 225 milliseconds, allowing it to react almost immediately when conditions change. This rapid adaptation means that operators can continue to drive or control a system safely, even when facing disruptions the vehicle has never encountered before. 

The researchers demonstrated the approach on real tracked vehicles and in simulation, including scenarios involving slippage, wind and uneven terrain. The system was also tested during the DARPA‑funded Learning Introspective Control (LINC) challenge, which focuses on helping vehicles cope with unforeseen conditions in real-world environments. 

Dr Maxime Allard, from Imperial’s Department of Computing, said: “Human operators are often placed in extremely challenging situations, where unexpected damage or environmental forces can quickly make a vehicle feel unpredictable or unsafe to control. We saw the consequences of this during the Ever Given incident in the Suez Canal, where wind and hydrodynamic forces overwhelmed the ship’s handling."

“FLAIR is designed to support operators in exactly these moments — when a vehicle is still functional, but no longer behaves as expected. By learning online, in milliseconds, how these disturbances affect the system, FLAIR helps restore intuitive control without requiring new controllers or lengthy retraining.” 

In experimental trials, vehicles using FLAIR recovered up to 74 per cent of normal operability after disruption — around twice as much as existing optimal and adaptive control methods. An online deep reinforcement learning approach was unable to recover control safely, largely due to the limited data available during live operation. 

The researchers say the work represents an important step towards safer deployment of robots and intelligent vehicles in complex, unpredictable environments. 

By focusing on assisting human operators, rather than full autonomy, the approach could help bridge the gap between today’s controlled robotic systems and their future use in transport, logistics, infrastructure inspection and emergency response — where rapid human‑machine cooperation is essential. 

Fast Learning-Based Adaptation for Immediate Recovery (FLAIR)