Curious Now

Story

Robotic Agentic Platform for Intelligent Electric Vehicle Disassembly

Artificial IntelligenceMaterials & Engineering

Key takeaway

Researchers have developed a robot system to automatically disassemble electric vehicle batteries, which could make recycling these batteries faster and more efficient as electric cars become more common.

Read the paper

Quick Explainer

The RAPID system presents a modular robotic platform for disassembling full-size electric vehicle battery packs. It combines a robotic arm, specialized end-effector, and computer vision to automate the process. The key steps are: 1) Identifying battery components like screws and busbars using an object detection model, 2) Planning an optimal disassembly sequence, and 3) Removing fasteners using a combination of pre-programmed poses, vision-guided execution, and visual servoing. This flexible approach is designed to handle the geometric and visual variability of diverse battery pack designs, going beyond prior work focused on smaller hybrid batteries.

Deep Dive

Technical Deep Dive: Robotic Agentic Platform for Intelligent Electric Vehicle Disassembly

Overview

This work presents a human-robot collaborative research platform called RAPID for disassembling full-size electric vehicle (EV) battery packs. The key contributions are:

  • A modular robotic system capable of unscrewing and removing battery components
  • An open-world vision pipeline for identifying screws, nuts, busbars, and other battery parts
  • Evaluation of fastener removal strategies including taught poses, vision-guided execution, and visual servoing
  • Integration of large language models (LLMs) for translating high-level disassembly instructions into structured robot actions

Problem & Context

  • Rapidly growing adoption of EVs creates an urgent need for scalable battery recycling
  • Manual disassembly of EV battery packs is laborious, dangerous, and uneconomical due to high design variability
  • Existing robotic disassembly approaches focus on smaller hybrid batteries, not full-size EV packs
  • Flexible automation is needed to handle diverse battery geometries and visual appearances with minimal training

Methodology

System Overview

  • The RAPID system integrates a gantry-mounted UR16e robot arm, a custom nut-running end-effector, and an RGB-D camera
  • It is designed to handle the full-size 800V, 77.4kWh Hyundai Ioniq5 EV battery

Disassembly Workflow

  1. Manually label battery components (screws, nuts, busbars, etc.) and store in a JSON representation
  2. Use perception pipeline to detect and localize labeled components during execution
  3. Plan an optimal disassembly sequence based on a traveling salesman problem formulation
  4. Execute removal of fasteners using taught poses, vision-guided execution, or visual servoing

Perception

  • Finetune YoloWorld object detector on a manually labeled dataset
  • Store 3D detections in a kD-tree to enable efficient lookup and merging of new detections

Fastener Removal

  • Use force control to apply downward force and engage the nut-runner
  • Compare taught poses, vision-guided execution, and visual servoing strategies
  • Evaluate impact of nut-runner extension length on success rate

Agentic AI Integration

  • Expose robot functionality via ROS-based Model Context Protocol (MCP) or as explicit SmolAgents tools
  • Evaluate performance of GPT-4o-mini and Qwen 3.5 LLMs on edge hardware

Results

Planning

  • Closed-form IK solver outperforms iterative KDL on success rate and planning time
  • Constrained RRT reduces end-effector travel but increases planning time

Vision

  • YoloWorld open-world detector achieves 97.6% mAP50, 76.3% mAP50-95
  • Comparable or better performance than conventional class-based Yolov11L

Fastener Removal

  • Taught poses: 97% success rate, 24 min duration
  • Vision-guided execution: 57% success rate, 29 min duration
  • Visual servoing: 83% success rate, 36 min duration
  • Longer nut-runner extension improves success rate from 47% to 97%

Agentic AI

  • Tool-based interface outperforms ROS MCP in reliability and efficiency
  • MCP failures often involve premature task completion without physical execution

Interpretation

  • Robots operate slower than humans (22 min vs. 17 min) but achieve high success rates (97%)
  • Vision struggles with reflective surfaces, leading to depth estimation errors in visual servoing
  • Agentic AI performance improves dramatically when using explicit, structured tool interfaces
  • Selecting the right task abstractions is critical for reliable LLM-driven robot control

Limitations & Uncertainties

  • Taught poses may not be robust to variations in battery registration
  • Visual servoing has difficulty converging due to poor depth estimates
  • Failure modes include missing screws, duplicate detections, and premature task completion
  • Scalability to full disassembly workflows with human-robot collaboration remains an open challenge

What Comes Next

  • Extend the platform to perform complete disassembly sequences with human assistance
  • Incorporate mobile manipulation to enable flexible tool exchange and part transport
  • Enhance sensor fusion for more reliable depth estimation and fastener detection
  • Further develop the agentic AI framework to handle complex error conditions and anomalies

Sources:

Source