EV Battery Disassembly: A Smarter Approach with Robotics

Author: Denis Avetisyan

Researchers have developed a collaborative robotic platform that uses artificial intelligence to improve the efficiency and precision of electric vehicle battery disassembly.

The convergence of established automation frameworks with emergent agentic AI, open-world vision, and advanced motion planning suggests a shift toward systems capable of navigating complexity-a synthesis where resilience isn’t measured by longevity, but by adaptive capacity within inevitable decay.

This work details an agentic robotic system leveraging ROS2, visual servoing, and large language models for automated EV battery disassembly with human collaboration and explicit tool definitions.

The increasing volume of electric vehicle batteries presents a significant recycling challenge, yet current disassembly processes remain largely manual due to design variability. This paper introduces the ‘Robotic Agentic Platform for Intelligent Electric Vehicle Disassembly’-a system designed to investigate perception-driven manipulation and AI-assisted robot programming for realistic battery recycling scenarios. Experimental results demonstrate improved fastener removal strategies-achieving up to 97% success rates-and highlight the benefits of structured tool interfaces for large language model-driven robotic control. Can this platform pave the way for fully autonomous, scalable robotic workflows that address the urgent need for sustainable battery recycling solutions?

The Inevitable Cascade: Addressing the EV Battery Recycling Challenge

The escalating adoption of electric vehicles is creating a rapidly expanding surge in end-of-life batteries, presenting a formidable recycling challenge for industries worldwide. These batteries, containing valuable materials like lithium, nickel, and cobalt, cannot simply be discarded; however, current recycling processes are bottlenecked by the need for efficient disassembly. Extracting reusable components and materials requires separating complex battery packs, a task traditionally performed manually. This approach is not only slow and costly but also carries inherent safety risks due to the high voltages and potentially hazardous chemicals involved. Consequently, the development of automated and streamlined disassembly methods is critical to unlock a sustainable circular economy for electric vehicle batteries and prevent a growing environmental burden.

The current reliance on manual disassembly for end-of-life electric vehicle batteries presents substantial obstacles to a truly sustainable battery lifecycle. This process is inherently slow and requires significant labor investment, creating a bottleneck in the growing flow of retired batteries. Beyond the economic inefficiencies, manual disassembly introduces considerable safety risks to workers, stemming from the high voltages and potentially hazardous materials contained within the battery packs. These risks necessitate specialized training and protective equipment, further increasing costs and complexity. Consequently, the limitations of manual approaches impede the large-scale recovery of valuable materials – like lithium, nickel, and cobalt – hindering the development of a closed-loop battery supply chain and delaying widespread adoption of truly circular economy principles within the electric vehicle industry.

The RAPID system is demonstrated successfully dismantling a Hyundai Ioniq5 electric vehicle battery.

Orchestrated Disassembly: Introducing the RAPID System

The RAPID disassembly system is a research platform designed for automated electric vehicle (EV) battery pack disassembly. It integrates a Universal Robots UR16e collaborative robot, a linear gantry system providing expanded workspace, and a dedicated nut-running tool. The UR16e robot manipulates battery components, while the linear gantry enables access to larger battery packs and facilitates multi-axis movement. The nut-runner is specifically incorporated to safely and efficiently remove fasteners securing battery modules, enabling component separation for inspection, repair, or material recovery.

The RAPID disassembly system achieves efficient and safe electric vehicle (EV) battery component separation through the synergistic combination of its hardware. The UR16e robot provides six degrees of freedom for accessing and manipulating battery modules within the assembly. A linear gantry extends the workspace, enabling disassembly of larger battery packs and accommodating varying component layouts. Crucially, the integrated nut-runner delivers precisely controlled torque to fasteners, preventing damage to battery cells and ensuring complete, yet gentle, separation of components. This combination of reach, dexterity, and controlled force is essential for automating a process that traditionally requires significant manual effort and poses potential safety risks.

The RAPID disassembly system incorporates human-robot collaboration to improve workflow efficiency. Specifically, human operators manage tasks requiring complex problem-solving, adaptability to unexpected variances in battery module condition, and final quality checks. The UR16e robot, with its integrated force-torque sensor, handles repetitive and physically demanding operations such as fastener removal and component separation, maintaining consistent precision and reducing the risk of damage. This division of labor leverages human cognitive abilities alongside robotic strength and repeatability, resulting in a disassembly process that is both faster and more reliable than either approach employed in isolation.

An industrial workflow utilizes agentic AI to coordinate a human co-worker in battery disassembly, beginning with part labeling and training, followed by a scan and registration process, and relies on a JSON-based description of parts and sequence for execution.

Intelligent Choreography: Agentic AI and the SmolAgents Framework

SmolAgents is an agentic AI framework designed to bridge the gap between large language models (LLMs) and robotic actuation for high-level task planning. This framework enables LLMs to not simply describe actions, but to actively control robotic systems. It functions by translating natural language instructions into a series of executable robotic commands. The agentic approach allows for dynamic planning and adaptation, as the LLM can reason about task requirements and generate sequences of actions based on the current state of the environment and available robotic capabilities. This contrasts with traditional robotic control systems that rely on pre-programmed sequences or direct human intervention, offering increased autonomy and flexibility in complex tasks.

SmolAgents achieve operational flexibility through the implementation of a Model Context Protocol (MCP). The MCP functions as a service discovery mechanism, allowing agents to dynamically identify and utilize available robotic capabilities without requiring pre-programmed knowledge of specific hardware or software configurations. This protocol facilitates communication between the agent and various robot services, detailing available actions, required inputs, and expected outputs. By decoupling the agentic AI from the underlying robotic infrastructure, the MCP enables the system to adapt to changes in hardware, software updates, or the introduction of new robotic tools without requiring modifications to the core agent logic. This adaptability is crucial for operating in dynamic environments and scaling the system to incorporate a broader range of robotic functionalities.

Disassembly sequences are optimized by formulating the task as an instance of the Traveling Salesman Problem (TSP). This involves defining each battery component as a “city” and the time required for the robot to navigate and manipulate between components as the “distance” between cities. Applying the TSP algorithm allows the system to determine the shortest possible path for disassembly, minimizing the total cycle time. This optimization is critical for maximizing the efficiency of the battery disassembly process, reducing operational costs, and increasing throughput by strategically ordering the removal of components.

Battery disassembly tasks are executed with guidance from an object detection system utilizing YoloWorld. This system identifies key battery components with a reported recall rate of 97%. The identified components then inform robotic actions, enabling the system to locate and manipulate specific parts during the disassembly process. Performance was evaluated based on the ability to correctly identify components necessary for disassembly, and the 97% recall indicates a high degree of accuracy in component localization for robotic manipulation.

Despite visually similar performance metrics-represented as quartiles, interquartile range, and outliers across ten trials-the model-controlled policy (MCP, orange) exhibited a 43.3% failure rate in the complex screw-removal task, while the explicit SmolAgents tool (blue) completed the task successfully every time.

Precision and Controlled Demolition: The Art of Screw Removal

The robotic system utilizes force control and visual servoing to reliably remove screws from battery components without causing damage. Force control involves actively monitoring and regulating the force applied by the robotic end-effector, preventing excessive torque that could strip screw heads or harm surrounding materials. Simultaneously, visual servoing employs camera feedback to precisely align the removal tool with the screw, compensating for minor positional inaccuracies and ensuring proper engagement. This combined approach allows the robot to adapt to variations in screw placement and component tolerances, maintaining consistent and controlled removal actions throughout the process.

The robotic system utilizes force control and visual servoing to compensate for inherent variability in battery component placement during screw removal. Force control measures the interaction force between the tool and the screw, allowing the robot to adjust its approach and prevent stripping or damage, even if the screw is not perfectly aligned. Simultaneously, visual servoing employs camera feedback to dynamically correct tool positioning in real-time, accounting for minor positional deviations in the battery pack assembly. This combined approach enables consistent application of the appropriate removal force and accurate tool alignment, irrespective of small-scale variations in component positioning that may occur during production or disassembly.

The RAPID system demonstrates a 97% success rate in screw removal operations, a performance level achieved through the combined implementation of force control, visual servoing, and ROS2 integration. This figure represents the percentage of screws successfully removed from battery components during testing, consistently meeting or exceeding established performance benchmarks. Data indicates this success rate is maintained even with minor variations in component positioning and screw tightness, highlighting the robustness of the combined approach. The metric is calculated based on a statistically significant sample size of removal attempts across multiple battery assemblies.

The Robot Automation Platform for Integrated Disassembly (RAPID) system utilizes the Robot Operating System 2 (ROS2) middleware to enable communication and control between its constituent robotic components. ROS2 provides a publish-subscribe messaging framework, allowing for real-time data exchange between hardware such as robotic arms, vision systems, and force sensors. This distributed communication architecture facilitates coordinated motion planning, sensor data fusion, and fault tolerance. Specifically, ROS2’s Data Distribution Service (DDS) ensures reliable and low-latency communication crucial for precise screw removal, and its modular design allows for easy integration of new sensors or actuators as the system evolves. The use of ROS2 also standardizes the software interface, promoting code reusability and simplifying system maintenance.

A custom end-effector integrating a nut-runner-retrofitted with an Arduino interface for Ethernet-based speed control and current measurement-and an Intel RealSense D435 provides a complete system for automated fastening as depicted in the accompanying block diagram.

Towards a Sustainable Cycle: Scaling Versatile Battery Recycling

The evolving landscape of electric vehicle batteries demands recycling solutions capable of accommodating diverse chemistries and voltages, and this automated system rises to that challenge. Unlike many current processes optimized for single battery types, this technology is engineered for versatility, effectively processing both high-voltage 800V systems and batteries utilizing Lithium Nickel Cobalt Oxide (LNC) as well as Lithium Iron Phosphate (LFP) chemistries. This adaptability is crucial, as the market features an increasing mix of battery technologies, each requiring specific handling protocols for safe and efficient disassembly. By avoiding the need for retooling or process adjustments based on battery type, the system streamlines the recycling workflow and enhances its overall scalability for future demands.

Automated battery disassembly offers a substantial pathway to enhanced recycling efficiency and diminished dependence on labor-intensive processes. Traditional EV battery recycling often relies heavily on manual disassembly, a method that is both time-consuming and susceptible to inconsistencies. This new approach utilizes robotic systems to carefully separate battery components – including cells, modules, and thermal management systems – with greater speed and precision. The result is a streamlined workflow capable of processing a higher volume of batteries, reducing operational costs, and improving the recovery of valuable materials like lithium, nickel, and cobalt. By minimizing human intervention, the system also addresses safety concerns associated with handling potentially damaged or unstable battery components, contributing to a more sustainable and economically viable recycling infrastructure.

Recent advancements in automated battery disassembly demonstrate a speed comparable to skilled human workers, completing the process in 22 minutes. This timing is remarkably close to the 17 minutes typically required for a trained technician to manually deconstruct an electric vehicle battery pack. The near-equivalence in speed, coupled with the potential for higher throughput and reduced labor costs, positions this automated system as a viable solution for scaling battery recycling operations. Such efficiency is crucial for meeting the growing demand for recovered battery materials and establishing a truly circular lifecycle for electric vehicle components, minimizing waste and maximizing resource utilization.

The development of automated battery disassembly technologies promises a significant leap towards a truly circular economy for electric vehicle batteries. Currently, end-of-life EV batteries represent a substantial waste stream and a missed opportunity for resource recovery; however, efficient automation facilitates the retrieval of valuable materials like lithium, nickel, and cobalt. By closing the loop on these critical resources, the need for environmentally damaging mining operations is reduced, and reliance on volatile supply chains is lessened. This streamlined process not only minimizes the environmental footprint associated with battery production and disposal, but also establishes a sustainable system where materials are continuously reused, fostering a more resilient and responsible approach to energy storage and consumption.

The presented robotic platform, designed for electric vehicle battery disassembly, inherently acknowledges the inevitable entropy of any complex system. The pursuit of automated disassembly, leveraging agentic AI and precise tool definitions, isn’t about achieving perpetual functionality, but rather about managing the system’s decline with increased efficiency. As Ken Thompson observed, “Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.” This resonates with the challenges of building such a system; anticipating all failure modes and edge cases-the ‘debugging’ of a physical, collaborative process-demands a humility regarding the system’s inherent imperfections and the constant need for adaptation and refinement. The platform’s success isn’t measured solely by its initial performance, but by its capacity to learn and evolve through iterative improvements, gracefully accepting the inevitable march of time and the accumulation of incidents as steps toward maturity.

What Lies Ahead?

This work, concerning the collaborative disassembly of electric vehicle batteries, represents a necessary, if predictable, step toward automating processes initially conceived for human dexterity. The platform’s success hinges on explicit tool definition, a constraint highlighting a fundamental truth: intelligence often manifests as skillful limitation, not boundless generality. Every bug in the system, every instance of misidentification or failed manipulation, is a moment of truth in the timeline, a record of the gap between intention and execution.

The reliance on agentic AI, while promising, introduces a new form of technical debt. The ‘past’s mortgage’ is now paid not in lines of code, but in accumulated training data and the implicit biases embedded within large language models. As these models evolve, the platform’s behavior will subtly shift, demanding continuous recalibration. The true measure of this research will not be its initial performance, but its capacity to gracefully age-to adapt, refine, and maintain functionality in the face of inevitable entropy.

Future iterations should address the inherent limitations of visual servoing in unstructured environments. While the current system excels with defined tools, the chaotic reality of end-of-life vehicle components presents a formidable challenge. The long-term viability of automated disassembly rests not solely on increasingly sophisticated algorithms, but on the development of robotic systems capable of robustly handling imperfection – of accepting that, in the realm of physical systems, complete control is an illusion.

Original article: https://arxiv.org/pdf/2603.18520.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/