The emerging circular economy introduces new challenges and new application areas for automation. One of them is EV battery disassembly. Unlike traditional industrial environments, battery disassembly operates without stable geometries, consistent designs, or reliable data.
Scaling Battery Disassembly: Why the Real Challenge Is Not Automation
Article from | Circular Battery Cluster
Automation has long been perceived as a magic pill for industries requiring scaling and production increase. In fact, it presents many benefits – it can make repeatable operations faster and more efficient, reduce reliance on manual labour, and minimise operational costs. Today, AI further enhances these capabilities. It is no longer a fantasy that future work will be a partnership between people, AI systems, and robots.
But not every industry is the same. The emerging circular economy introduces new challenges and new application areas where automation is needed. One of them is EV battery disassembly. Unlike traditional industrial environments, battery disassembly operates without stable geometries, consistent designs, or reliable data.
This shifts the problem from execution to adaptation: not how to automate a process, but how to make automation work in a system where no two inputs are the same. Several European projects funded under Horizon Europe — including RECIRCULATE, REBELION, REINFORCE, and BATTEREVERSE — are actively working on this challenge. We spoke with experts from the University of Birmingham, Centria University of Applied Sciences, Aston University, Comau, and Eurecat Technology Centre to gather their perspectives on recent technological advancements, as well as the key progress made and challenges that remain.
The key challenges are still there
The key challenges in automating battery disassembly do not stem from a lack of technology, but from the nature of the system itself. Across multiple industry and research initiatives, experts consistently point to extreme variability in battery design as the primary barrier. Differences in geometries, fastening systems, and internal architectures, combined with the lack of standardisation, make it difficult to generalise robotic processes across battery types.
As Alireza Rastegarpanah, Assistant Professor in Robotics and CoPI of REBELION project, explains, “one of the main challenges is the diversity in battery design and the lack of standards makes robotizing the process much more difficult.”
This is why the field is increasingly moving towards operation in unconstrained environments, where systems must adapt to unknown inputs rather than rely on predefined setups. As Rustam Stolkin, Professor, Chair of Robotics, University of Birmingham, Founder and Director of the Extreme Robotics Lab, notes, “we are moving towards systems that can function in unconstrained environments, where the robot has to deal with unknown objects and conditions.”
This variability directly impacts scalability. Instead of repeatable processes, each new battery often requires additional programming, validation, and engineering effort. As Giampiero Pupillo, Automation Engineer, Comau representing the REINFORCE project, puts it, “traditional automation works extremely well in repeatable environments, but battery dismantling is the opposite: there is no stable reference, no guaranteed geometry, no consistency.” Recycling facilities operate under inherently unstable conditions, processing a heterogeneous mix of batteries, often damaged or deformed and without reliable prior data, which breaks the core assumptions on which traditional automation is built.
Beyond system-level variability, materials and component design introduce additional layers of complexity. Certain elements remain inherently difficult for robotic handling. As Tomi Pitkäaho, PhD, Principal Lecturer in Research, Centria University of Applied Sciences, notes, “one of the most challenging tasks is removing the wiring harness. It’s highly deformable — and that makes it very difficult for robots to handle.” Flexible components, unlike rigid parts, break the assumptions of predictable manipulation and require far more advanced sensing and control strategies. Material choices further complicate the process: adhesives (widely used in battery design) create significant barriers to disassembly, sometimes requiring unconventional approaches such as freezing bonded components with liquid nitrogen.
Another critical limitation is the lack of structured and accessible battery data. In many cases, teams are forced to rely on indirect sensing methods, such as combining RGB and thermal imaging, or even synthetic datasets, to compensate for missing real-world information. At the same time, efforts are underway to build component-level datasets that can support both automation and traceability. As Alireza Rastegarpanah highlights, such datasets are essential not only for vision-guided robotics, but also for enabling digital battery passports and tracking batteries across their lifecycle from production to end-of-life and potential second use. However, as Óscar Palacín, Researcher, R+D+i Project Manager, Eurecat – Technology Centre, points out, the current lack of high-quality data significantly limits these efforts and reflects how early the field still is.
Another critical challenge highlighted across all projects is data availability. In many cases, access to reliable battery data, in particular, SoH, remains limited. As a result, if the condition of a battery cannot be determined, it is typically directed to recycling rather than reuse, regardless of its remaining potential. This makes data not just a technical enabler, but a key bottleneck that is still difficult to address in practice.
These challenges are not new. From the very beginning, it has been clear that battery disassembly is inherently complex: batteries vary significantly in design, involve hazardous and sensitive materials, and were never manufactured with disassembly in mind. These issues have been consistently highlighted across research and industry and they remain not fully resolved. However, this does not mean progress has stalled. On the contrary, research is actively advancing, and recent projects demonstrate meaningful results.
What automation can already deliver
Despite all abovementioned constraints, progress is tangible.
The RECIRCULATE project develops an AI-based automated dismantling system aimed at achieving over 98% classification accuracy across 20 different battery types and detecting 10 defect classes with over 99% classification accuracy, while meeting strict time targets (under 60 minutes for pack-to-module and 120 minutes for module-to-cell disassembly).
As a result, Centria University of Applied Sciences developed a system built around a KUKA KR 10 industrial robot mounted on a linear unit, equipped with multiple end effectors for different tasks (bolt gun for screws, vacuum gripper for lids and plates, and a Schunk gripper with custom jaws for cables and connectors). The system relies on six machine learning models, including battery identification, defect detection (e.g. punctures, dents, thermal damage, missing components), and component recognition.
A key advantage of such systems is their ability to reuse knowledge.
“If we can identify the battery and recognize that we’ve disassembled this type before, we can reuse the same disassembly strategy,” Tomi Pitkäaho says.
The system presents a hybrid approach: the pack-to-module stage is AI-driven, enabling identification and reuse of disassembly strategies for known battery types, while the module-to-cell stage remains fixture-based.
Overall, the project demonstrates a working proof of concept: high-accuracy, AI-supported battery disassembly is achievable through a combination of robotics, machine learning and structured system design.
Within the REBELION project, the University of Birmingham leads the automation and robotics work, focusing on flexible, research-driven (low-TRL) solutions, while partners such as UPV work on more practical, higher-TRL implementations, including partially automated dismantling of light EV batteries combining manual operations with vision-guided robotics.
As mentioned by Rustam Stolkin, the key focus of Birmingham’s work is enabling robots to operate across highly variable and unknown battery types. To address this, the team is developing an adaptive digital twin, integrating real-time data from vision systems, simulations, and physical robots to plan, synchronise, and execute complex operations. The system supports multiple modes of interaction from full autonomy to VR/AR-based teleoperation and human-in-the-loop control, including haptic feedback that allows operators to “feel” the task remotely in hazardous environments.
On the perception and manipulation side, REBELION focuses on battery and component recognition in cluttered environments and generalisable grasping of unknown objects. Notably, instead of relying on neural networks, the system uses geometric methods, enabling fast and robust handling of previously unseen components. The project also develops autonomous cutting capabilities, where a robot builds a real-time 3D model, plans and simulates a cutting path, and executes it under human supervision.
Further work includes shared control systems, where AI suggests optimal actions (e.g. grasp points) while the human operator retains control, as well as multi-robot coordination, allowing multiple robots to divide tasks and avoid collisions in real time. Recent experiments also explore the use of LLMs, enabling non-expert users to control robots via natural language commands.Finally, REBELION is scaling these approaches to heavy-duty robotic systems capable of handling large battery packs (up to ~500 kg). This introduces additional challenges related to safety, control, and system integration, which we are addressing as the project progresses.
The BATTEREVERSE project takes a full lifecycle approach to battery reverse processing, integrating data, automation, and AI into a unified system. It starts with capturing battery data to create a digital battery passport, including information on chemistry, condition, and usage history, followed by a combined process of discharge, SoH assessment, and safe deactivation, supported by secure packaging solutions.
From there, batteries are directed to either recycling or repurposing pathways. On the recycling side, the project focuses on semi-automated robotic dismantling and computer vision-based sorting, while repurposing relies on acoustic diagnostics and machine learning to assess remaining useful life and identify second-life applications. All stages are connected through a digital twin, enabling simulation and optimisation of reverse logistics and operations.
A key component is collaborative robotic dismantling, where robots perform tasks such as unscrewing using vision systems and adaptive tools, with human–robot collaboration handling variability and exceptions. In parallel, a multimodal inspection system (RGB + thermal imaging) detects anomalies and ensures safety, while synthetic data generation is used to compensate for the lack of real defect data in training AI models.
Overall, the project demonstrates a scalable system architecture that combines robotics, AI-driven inspection, and data integration. However, it also reinforces a broader industry insight: while automation in battery disassembly and inspection is technically achievable, its scalability remains constrained by variability, leaving open the question of whether this challenge will be solved through advances in AI, system design, standardisation, or continued human involvement in the loop.
While these projects approach the problem from different angles — from system-level integration to adaptive robotics and lifecycle optimisation — they are ultimately addressing the same core challenge: how to make automation work under extreme variability. Together, they provide a 360-degree view of the problem, revealing that no single solution is sufficient, but that progress emerges from combining the most effective approaches across robotics, AI, and system design.
What’s next: Scalability, Hybrid Systems, AI and LLMs
Giampiero Pupillo argues that the focus on replacing manual work is somewhat misleading. “When we talk about automation in battery recycling, the immediate focus is on replacing manual operations. But that’s not the real challenge.”
According to him, the industry needs to approach the problem on two levels — hardware and software, with the software layer potentially playing a more critical role in enabling scalability. This is where low-code approaches become relevant. Comau’s approach is based on developing a minimal set of reusable building blocks that can be applied across tasks such as vision, handling, and screwing. However, implementation remains complex, as programming and commissioning must account for a wide range of external variables.
Building on this idea, Rustam Stolkin from the University of Birmingham describes a closely related concept of “action primitives.” In this framework, robots are equipped with generalized actions — such as unscrewing, cutting, grasping, or moving — that can be applied across different objects and environments. A vision system identifies task-specific elements (e.g. a screw at a given location) and maps them to these predefined actions.
By encoding such actions in a generalizable way, systems can reuse them across different battery types. This creates a foundation for a low-code layer, where high-level instructions — such as defining positions for unscrewing — can be used to adapt processes to new batteries without rewriting the entire program.
At the same time, advances in AI and data-driven approaches are opening new possibilities — but also exposing new limitations. In projects such as REBELION, large language models are being explored as task planners, enabling non-expert users to interact with robotic systems through natural language. However, as Alireza Rastegarpanah notes, these approaches remain computationally intensive and difficult to deploy in real-time closed-loop systems.
Similarly, while computer vision models (e.g. YOLO-based systems) are widely used for object detection and localisation, core manipulation tasks such as grasp planning and cutting are often still based on geometric methods rather than learned models. As Rustam Stolkin explains, this allows systems to operate reliably on unknown objects without requiring large training datasets. The question of data availability remains critical. While some datasets have been released, most remain limited in scope, and large-scale, high-quality datasets are still largely unavailable. Even access to operational battery data (e.g. BMS data) depends on manufacturer cooperation, as highlighted by Tomi Pitkäaho.
Research teams are increasingly moving away from the idea of fully autonomous systems as the ultimate goal. Instead, hybrid models are gaining traction. Rustam Stolkin describes a spectrum of approaches combining automation with human input: “Some tasks can be fully automated, but others are too complex or too variable. That’s why we explore systems where humans remain in the loop.”
These systems range from teleoperation to shared control, where humans guide or intervene in robotic processes when needed. Across projects, a similar pattern emerges: scalability is not achieved by eliminating the human role, but by redefining it within hybrid environments that combine robotics, AI, and human judgment.
The implication is broader than a technical choice. In highly variable systems such as battery disassembly, full automation may not only be difficult — it may not be necessary. A significant share of operations can already be simplified and automated today, while others remain inherently non-scalable or inefficient to automate. The strategic focus, therefore, shifts to scaling what scales, and designing systems that intelligently combine automation with human intervention where it adds the most value.
Summary
Automation in battery disassembly is no longer a question of technical feasibility — it is already happening, as demonstrated by several European projects. The real challenge lies in scaling it under conditions of uncertainty, variability, and incomplete data.
As long as battery designs remain non-standardised, automation cannot rely on repetition alone. Instead, it must be built around flexibility combining modular software, AI-driven perception, and human-in-the-loop approaches. This shifts the paradigm: the goal is not to automate everything, but to scale what scales and integrate human judgment where it remains essential.
The industry is therefore not moving toward full autonomy, but toward adaptive, hybrid systems. In this context, the key question is no longer whether automation is possible, but how to design systems that can scale in environments that resist standardization by design.
The content & opinions in this article are the author’s and do not necessarily represent the views of RoboticsTomorrow
Featured Product
