End-of-life EV batteries and electronics come back in hundreds of variants, often damaged, and are still dismantled largely by hand under high-voltage risk. Vision-guided robots recognise the product variant, locate screws, connectors and cables, and perform contact-rich removal steps autonomously, with humans handling exceptions and safety release. The aim is safer, more scalable pack-to-module-to-cell processing that feeds reuse, second life and recycling.
Problem & value
- Today: pack dismantling is manual or semi-automated with fixed programmes per model. Workers face high-voltage and thermal-runaway hazards; throughput is low; variant diversity breaks classic teach-in robot programmes.
- AI value: perception makes a cell variant-tolerant (unknown or damaged products), reduces human exposure to live packs, and creates per-pack traceability that supports reuse decisions.
- Regulatory pull: the EU Battery Regulation (EU) 2023/1542 sets collection, material-recovery and recycled-content requirements and introduces the battery passport, which raises volumes of recycling, remanufacturing and second-life flows and rewards clean separation of modules and cells [3].
- Evidence: the EU-funded Recirculate project built a cell around a KUKA KR10 on a mobile linear track with a depth camera and custom tools; ML models identify screws, connectors and wiring, and the system separates packs into modules and modules into cells. Reported pack identification was near-perfect for Ford and Tesla packs, and the team reported about 18 months for this first phase; scaling to more battery types needs more training data [1][2]. The German research project AuBa@C (RWTH Aachen) pursues the same goal; no primary source was re-verified for this file.
- No sourced throughput or cost-per-pack figures are cited; most results are pilots.
Variants
- EV traction battery packs: lid removal, busbar/connector/cable removal, module extraction, module-to-cell separation (adhesives and welds are the hardest steps).
- Remanufacturing: electric motors, power electronics, compressors, gearboxes; non-destructive disassembly for reuse of intact parts.
- Consumer electronics: phones, laptops and e-bikes, with high variance and small fasteners.
- Second-life triage: disassembly to module level, combined with state-of-health testing (PRED-026).
How it works
- incoming pack (quarantine, discharge/safety check by qualified staff)
- identification (RGB-D + VLM/classifier, label/VIN/passport lookup)
- disassembly plan (variant database or generated from CAD / teardown graph)
- perception of fasteners, connectors, cables (detection + segmentation + 6D pose)
- tool selection (screwdriver, cutter, vacuum gripper, gripper) and force-controlled execution
- verification (camera + torque/force check; retry or replan)
- sorted output: modules/cells/fractions, traceability record
- human intervention on failure, damage or unknown variant
- Perception: open-vocabulary and fine-tuned detectors find screws and connectors; depth data gives poses; a classifier selects the stored programme for known variants.
- Planning: disassembly sequence planning on a precedence graph, partly learned from demonstration (ROBO-003); replanning when a step fails (rusted, deformed or missing fasteners).
- Control: impedance/force control for contact-rich tasks; learned policies are emerging but are not yet trusted for high-voltage work without classical guard logic.
- Human-in-the-loop: safety release, damaged-pack handling, teleoperated recovery, and approval of new variants.
Data preparation
Robotic disassembly depends on product knowledge more than on classic business data. Typical sources: manufacturer teardown instructions, CAD and service manuals (PDFs on file shares or in the DMS, sometimes paper), battery passport and BMS data where available, goods-in records in the ERP or WMS, and above all your own images: RGB-D scans of incoming packs, including damaged and aged ones. Typical gaps: no consistent pack identification (unreadable type labels, no mapping to variants), unstructured manuals, no labels for fasteners and connectors, and very few examples of rare variants and damage.
Preparation steps: build a variant catalogue keyed by pack ID with metadata (PREP-010), digitise and OCR manuals (PREP-006), capture goods-in images at a fixed station with consistent lighting, label fastener and connector locations with domain experts and reserve a test set (PREP-018), and harvest labels from operator corrections (PREP-021). Synthetic data from simulation (ROBO-010) fills gaps. Useful tools: CVAT or Label Studio for annotation, Open3D for point clouds, Paperless-ngx for manuals. Plan roughly half of the total effort for data work; it never ends completely, because new pack designs keep arriving. See also GUIDE-003.
Implementation paths
Which path for whom
- Enterprise: C, because disassembly know-how is core to a recycler or OEM, the perception stack must run locally in real time and variant data is a competitive asset; integrators (A) only fit a few well-known pack families.
- Developer: C, because ROS 2, open detectors and simulation give full control over perception and policies for research and product development.
- Public: B, because research institutes and municipal recyclers can combine an integrator cell with EU-hosted cloud services for off-line tasks and do not need to own the whole stack.
Path A — Buy
- Integrators and robot OEMs: KUKA (EU, Germany), ABB (EU/CH), Fanuc, Universal Robots, plus specialised EU recycling-automation integrators and recyclers with in-house cells. Verify scope: pack-level vs cell-level, supported pack list, high-voltage safety concept.
- Perception stacks: Cognex, Zivid and Photoneo (EU) 3D cameras, MVTec HALCON (EU) and Isra Vision (EU) software.
- Choose A when: you process one or a few well-known pack families at volume.
Path B — Build on cloud APIs
- Use a cloud VLM only for off-line tasks (variant identification from labels and photos, extracting teardown instructions from manuals, assisting operators); keep real-time perception and control local. Orchestrate via ROS 2 and MES/WMS integration with EU-region hosting.
Path C — Self-host (open source) (recommended for enterprise and developers)
- Stack: ROS 2 with MoveIt 2, open detectors (YOLO-family, RT-DETR, Grounding DINO, SAM 2), Open3D / PCL for point clouds, NVIDIA Isaac Sim / Isaac Lab or MuJoCo for simulation (ROBO-010), LeRobot for demonstration-based policies.
- Hardware: HW1 edge GPU (Jetson Orin class) per cell for inference; HW2/HW3 for training.
Path D — Custom training
Needed in practice: per-variant datasets of fasteners, connectors and cabling, including damaged and aged examples. Combine synthetic data from simulation with real teardown images; correction data from operator interventions feeds retraining. Expect an ongoing data effort as new pack designs arrive.
State of the art (2026)
- Pilot cells demonstrate pack-to-module-to-cell sequences with ML-based fastener and connector recognition [1][2]; variant coverage remains the bottleneck.
- Trends: vision-language-action models and demonstration learning (ROBO-004, ROBO-003) are explored for tool use; simulation and digital twins help train perception and plan sequences.
- Limits and failure modes: glued/welded cell stacks, corroded or deformed fasteners, swollen or damaged cells, hidden connectors, and unknown pack designs; thermal-runaway risk means the cell needs fire detection, inert or shielded design and conservative stop behaviour. Design-for- disassembly by OEMs would help more than any perception model.
Shared resources & combinations
- Same stack: ROBO-001 (bin picking: 3D vision and grasping), ROBO-003 (programming by demonstration), ROBO-010 (simulation for sim-to-real), ROBO-016 (AI safety functions for person detection around the cell).
- Upstream/downstream: PRED-026 (battery state-of-health for reuse decisions), VIS-020 (material stream recognition after shredding), DATA-021 (battery passport and product carbon footprint data).
Effort, cost & KPIs
- Path A: 6-12 months for specification, safety concept and commissioning of a cell.
- Path C/D: 12-24+ months for a robust multi-variant system; cost drivers are data collection, tooling, safety engineering and certification.
- KPIs: variants handled without manual programming, steps completed autonomously (%), intervention rate, cycle time per pack, module/cell damage rate, safety incidents, recovered material purity, share of components reused.
People & change
Start with management and the head of the recycling line (process owner) to fix scope: which pack families, which steps stay manual. Involve the operators who disassemble by hand early; they know the real variants and hazards and become the annotators, testers and later cell supervisors. IT-security covers network separation of the cell and remote access; legal/compliance handles machinery safety, high-voltage rules and liability; data owners release CAD, manuals and goods-in records.
Works council: yes. Robot cells log operator interventions and presence, which is objectively suited to monitor behaviour or performance, so §87(1) no. 6 BetrVG applies (and workplace safety co-determination); §90(1) no. 3 requires early information about planned AI-supported work processes, and §80(3) allows an expert (GUIDE-004). A DPIA is usually not needed because the product data is not personal, but camera views of staff must be checked with the DPO.
Support AI literacy (Art. 4 AI Act) with hands-on training for cell supervisors on limits, failure modes and stop procedures (GUIDE-006). Typical concern: job loss. Address it openly: the robot takes the hazardous high-voltage steps, staff move to supervision, exception handling and quality. Start with a pilot cell and one pack family, define go/no-go KPIs (GUIDE-009), and keep humans in control with emergency stop, manual override and approval of every new variant.
Risks, regulation & compliance
- EU AI Act: classification depends. Vision functions inside a machinery safety function are safety components of machinery (Annex I via the Machinery Regulation), which makes them high-risk if third-party conformity assessment applies [4][5]. A pure productivity perception stack outside safety functions is typically minimal risk. No Art. 5 issue is expected; keep cameras off workers' emotion or behaviour analysis.
- Machinery: Machinery Regulation (EU) 2023/1230 applies from January 2027 and covers self-evolving behaviour and cybersecurity [5]; ISO 10218 and ISO/TS 15066 apply to cobot use.
- Battery/waste law: Battery Regulation [3], Waste Framework Directive, ADR transport of damaged batteries, permits for hazardous-waste treatment, and electrical-safety rules for work on live systems (qualified persons).
- GDPR: limited; only camera coverage of workers needs a works-council agreement and privacy review.
- Mitigations: discharge and isolation before robot contact, independent safety PLC, fire detection and suppression, conservative stop logic, human release for each new variant.
Getting started
- Define the product scope: which pack families, volumes and target outputs (modules, cells, parts).
- Run a teardown study with manual disassembly timing and record video and RGB-D data of each step.
- Check Battery Regulation duties and the high-voltage safety concept with your safety officer.
- Pilot a perception-only stage (fastener/connector detection) with human execution, then add robotics.
- Add simulation (ROBO-010) and operator-correction logging to extend variants.
Sources
- Control.com - Recirculate disassembly system gives new life to old EV batteries - https://control.com/news/recirculate-disassembly-system-gives-new-life-to-old-ev-batteries/
- electrive - Recirculate tests AI-driven robotic system for EV battery disassembly (2 Sep 2025) - https://www.electrive.com/2025/09/02/recirculate-tests-ai-driven-robotic-system-for-ev-battery-disassembly
- Regulation (EU) 2023/1542 (Batteries Regulation) - https://eur-lex.europa.eu/eli/reg/2023/1542/oj
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2023/1230 (Machinery Regulation) - https://eur-lex.europa.eu/eli/reg/2023/1230/oj
- ROS 2 MoveIt - https://moveit.ai/