How AI helps take apart used electric-car batteries

An old electric-car battery is more than a container of raw materials. Some components may be candidates for reuse; others may need to enter a recovery process. Oak Ridge National Laboratory’s battery-disassembly work starts with that distinction. Diagnostics help determine whether to extract useful modules selectively or take the pack further apart. The purpose of opening it changes the work the robot should perform.[1]

AI can contribute at several points in that work: identifying components, choosing an action sequence and turning experience into a routine that can be repeated. These are distinct capabilities. A 2025 study of robotic bolt removal combines learned visual recognition with explicit planning rules, then investigates how language-model suggestions and stored trajectories might simplify repetitive operations.[4]

Recognising the part is the first handover

One useful vision approach is YOLO-World, presented at CVPR 2024. It learns relationships between regions in images and descriptions of objects. A user can supply the vocabulary that matters for a task, instead of being limited to the original detector’s fixed list of categories. The model can encode that vocabulary in advance, so it need not repeatedly process the same words for every new image. That makes language a way to configure the search.[3]

The handover is still precise: an object detector returns potential objects and their locations in an image. It does not, by that output alone, establish that a tool has engaged a fastener or that a component is free. A useful workshop system must connect the detected object with an operation and an observable result. A good label starts the job; it cannot certify that the job has finished.[3]

A full-size battery makes the trade-off visible

RAPID’s March 2026 preprint describes a gantry-mounted arm beside a Hyundai Ioniq 5 battery, combining YOLO-World-based vision with a nut-running tool. Researchers compared taught positions, locations from an initial scan and scan-based locations refined by further visual alignment close to the fastener.[2]

Teaching positions won this comparison, but it required a person to establish them. Additional visual alignment improved on the initial scan at the cost of time. Operators also corrected the detected inventory. Testing followed an earlier manual disassembly and reassembly. This was a prepared example of one full-size battery design, rather than a stream of unfamiliar retired packs. Full automated disassembly sequences remained future work.[2]

Experience can shorten the routine

The 2025 bolt study explored reusing movement patterns to avoid redundant recognition steps. After a repeated layout had been recognised, the robot could reuse a movement sequence instead of locating each bolt afresh. The resulting routines were faster in the laboratory, but positional errors could accumulate along the sequence. Tests used unobstructed M8 bolt arrangements under controlled lighting. The language model operated offline, proposing changes that became robot routines; it was not improvising each movement during removal. This is a way to improve a familiar repetitive task, with flexibility bounded by the layout it has learned.[4]

A connector needs a different kind of intelligence

Electrical connectors expose the limits of treating everything as a screw. In a study published in Robotics and Computer-Integrated Manufacturing, researchers designed a tool that first releases a connector’s latch and partly extracts the plug, creating enough exposed surface for a gripper. Force measurements help establish what is happening during contact. The mechanical sequence makes the problem easier before any more sophisticated decision-making is added.[5]

For alignment inside a battery, the team used reinforcement learning to tune parameters in a force-based search method. In 100 search experiments, the robot reached the target area within five steps. This was a result about finding the alignment area, not 100 complete batteries dismantled. The authors also limit the demonstrated approach to connectors in good condition; stuck or damaged parts remain further work. Learning the search and engineering the tool contribute different pieces of the solution.[5]

People can remain part of the physical process

Another research route keeps a person in direct control. A 2023 study used teleoperation for tasks on a Nissan Leaf battery module stack, including unbolting, removing a cover and separating modules. Operators controlled a remote robot through another robot or a haptic interface that conveyed forces. The arrangement lets the operator act through machinery while retaining judgement about contact and movement. It is human-controlled robotics, rather than autonomous AI.[6]

The comparison revealed why interface design matters. Reaching a component, aligning the tool, maintaining a grip and moving the part away all consume time. The faster arrangement did not deliver the higher success rate on every task. A disassembly station can therefore benefit from improvements to the operator’s view and control as well as from more capable perception or planning software.[6]

The useful output depends on the recovery plan

Fraunhofer’s DeMoBat project built separate demonstrators for distinct levels of disassembly, including extracting modules from packs and working from modules towards cells and materials. Its structure captures an important practical point: opening a case, retrieving an intact component and recovering material are different achievements. They require different tools and ways to assess success. A production process must connect those stages while preserving the value it is trying to recover.[7]

Oak Ridge’s account likewise describes testing battery condition before selecting a disassembly route. That reverses the instinct to make the robot remove everything it can reach. If the goal is selective reuse, unnecessary removal can add work without advancing the goal. If the goal is material recovery, stopping after opening the cover leaves much of the process unfinished. The destination should determine the sequence.[1]

The International Energy Agency’s 2025 battery outlook adds the economic context. It describes limited feedstock for recycling and complications for reuse businesses, including competition for retired batteries and falling prices of new ones. Efficient dismantling is therefore one part of a larger calculation. Recovering a component has to justify the work of accessing, checking and repurposing it; removal speed alone cannot answer that question. The report also stresses that limited available feedstock constrains how quickly recycling can reduce demand for newly mined materials. Better robots cannot recover batteries that have not yet reached that stream.[8]

This gives AI a useful, concrete target. A more adaptable robot can help make a chosen recovery route practical across parts it has to recognise and manipulate. The output worth pursuing is a component or material stream that can move to its next use. Counting the things an arm has successfully detached is a beginning; tracking what those things are fit for completes the picture.[7]

Report an error or suggest a correction