💻 computer science

Hybrid Task and Motion Planning with Reactive Collision Handling for Multi-Robot Disassembly of Complex Products: Application to EV Batteries

This paper presents a vision-driven, closed-loop Task-and-Motion Planning (TAMP) framework for multi-robot EV battery disassembly that integrates learning-based motion planning with a hybrid safety layer to significantly reduce path length, makespan, and collision risks compared to standard approaches.

Original authors: Abdelaziz Shaarawy, Cansu Erdogan, Rustam Stolkin, Alireza Rastegarpanah

Published 2026-04-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abdelaziz Shaarawy, Cansu Erdogan, Rustam Stolkin, Alireza Rastegarpanah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have two highly skilled robot chefs working together in a tiny, chaotic kitchen to take apart a complex, dangerous cake (an electric vehicle battery). The goal isn't just to eat the cake; it's to carefully dismantle it piece by piece without making a mess, hurting themselves, or blowing up the kitchen.

This paper describes a new "brain" for these robots that helps them do this job much better than before. Here is how it works, broken down into simple concepts:

1. The Problem: Two Chefs in a Crowded Kitchen

Previously, if you told two robots to work together, you might give them a fixed list of instructions: "Robot A, cut the top. Robot B, lift the side." This is like giving a script to actors who can't improvise.

The Issue: If Robot A moves too fast, Robot B might bump into it. If a piece of the cake falls in a weird spot, the robots get stuck because they were following a rigid script.
The Danger: EV batteries are heavy and dangerous. If the robots crash, it could be a fire hazard.

2. The Solution: A "Smart Brain" with Three Superpowers

The authors built a system that combines three different ways of thinking to make the robots smarter, safer, and faster.

A. The "Eagle Eye" (Vision & Re-scanning)

Instead of just looking at the kitchen once at the start, these robots keep looking around constantly.

The Analogy: Imagine playing a game of "Hot and Cold." Every time a robot picks up a piece of the battery, it stops, looks around again with its cameras, and asks, "Okay, what's left? Where did that piece land?"
Why it helps: If a piece moves or gets hidden, the robots don't get confused. They update their mental map instantly and change their plan on the fly.

B. The "Memory of a Master Chef" (Learning from Experience)

This is the coolest part. The robots don't just guess how to move; they "remember" how a human expert did it before.

The Analogy: Think of a jazz musician. A beginner might play random notes (this is what standard robots do). But a master musician has practiced specific riffs and knows exactly how to move their fingers to hit the perfect note.
How it works: The researchers showed the robots how to dismantle the battery by hand (or via teleoperation). The robots learned these "riffs" using a mathematical tool called a Gaussian Mixture Model (GMM). When it's time to move, the robot doesn't just search randomly for a path; it starts with the "expert's path" as a guide. This makes their movements much smoother and shorter, like a dancer who knows the choreography rather than a person flailing around.

C. The "Dance Partner" (Predictive + Reactive Safety)

The robots have two layers of safety, like a dancer who plans their steps but also reacts if their partner stumbles.

Predictive (The Plan): Before moving, they use a "digital twin" (a virtual copy of the kitchen) to check, "If I move here, will I hit Robot B?"
Reactive (The Reflex): While they are moving, if Robot B suddenly jerks or an obstacle appears, the robots have a "reflex." They instantly steer away, like a driver swerving to avoid a pothole, and then immediately calculate a new path.
The Result: They don't just hope they won't crash; they actively dodge and weave while staying on task.

3. The Results: A Much Better Performance

When they tested this new system on real EV batteries, the results were impressive:

Less Walking: The robots walked (moved their arms) 63% less distance. Instead of wandering aimlessly, they took the most direct, expert-like paths.
Faster Job: The whole job finished about 8% faster.
Safer Space: The robots occupied much less "air space" (swept volume). Imagine two people dancing in a small room; the new system made them dance so efficiently that they barely brushed against each other, whereas the old system had them bumping into furniture constantly.

The Big Picture

This paper isn't just about taking apart batteries. It's about teaching robots how to collaborate in messy, real-world situations.

Instead of being rigid machines that follow a script, these robots are now like a well-rehearsed dance troupe. They watch the stage, remember the choreography, and if someone trips, they instantly adjust their steps to keep the show going safely. This makes them ready for the future, where robots will need to work alongside humans in factories, hospitals, and homes, handling complex tasks without needing a human to hold their hand every second.

1. Problem Statement

The paper addresses the challenge of coordinating multiple robots for complex, sequential manipulation tasks in dynamic and cluttered environments. Specifically, it targets the disassembly of Electric Vehicle (EV) batteries, a safety-critical domain characterized by:

Strict Precedence: Tasks must follow a specific order (e.g., removing screws before modules).
Dynamic Uncertainty: Object positions may shift, and obstacles (including other robots) move during execution.
Dual-Arm Coordination: Two robots must operate in a shared workspace without colliding, requiring tight synchronization between task planning (what to do) and motion planning (how to move).
Limitations of Current Approaches: Traditional methods often decouple task and motion planning (open-loop), leading to failures when perception noise or dynamic obstacles invalidate pre-computed plans.

2. Methodology

The authors propose a perception-driven, closed-loop Task-and-Motion Planning (TAMP) framework implemented on a dual-robot platform (two UR10e cobots). The system integrates four main layers:

A. Perception and Task Layer

Sensing: Uses Intel RealSense RGB-D cameras with YOLOv8-Seg for real-time object detection, segmentation, and 3D pose estimation of battery components (cables, busbars, cells, plugs).
Task Decomposition & Allocation:
- An expert-encoded lookup table enforces safety-critical disassembly precedence and tool compatibility (e.g., 3-finger gripper for screws, vacuum gripper for modules).
- Closed-Loop Update: Unlike static planners, the system re-scans the scene after every action. It updates the completion status of objects and re-allocates remaining tasks dynamically based on the current scene state.

B. Motion Planning Layer (Predictive)

Learning-from-Demonstration (LfD): The system utilizes Task-Parameterized Gaussian Mixture Models (TP-GMM) trained on human expert demonstrations.
GMM-Informed RRT: The TP-GMM generates a high-probability region in the configuration space. This acts as a constraint for a Bi-RRT (RRTConnect) planner, biasing the sampling process toward expert-like, efficient trajectories rather than random exploration.
Digital Twin: A MoveIt/FCL (Flexible Collision Library) digital twin performs predictive collision checking before execution.

C. Reactive Execution Layer

Hybrid Safety Strategy: Combines predictive checking with reactive avoidance.
Collision Monitoring: Uses an OctoMap (0.03m resolution) updated at ~10Hz to detect dynamic obstacles. A safety threshold ( $\epsilon = 0.15$ m) triggers reactive responses.
Reactive Avoidance: If a near-collision is detected, a velocity vector steering strategy (using Rodrigues' rotation formula) reorients the robot's tool velocity to steer away from the obstacle.
Replanning: If avoidance fails or the scene changes significantly, the system triggers an immediate replan from the current configuration to the target.

3. Key Contributions

Hybrid Predictive/Reactive Collision Handling: A novel integration of digital-twin predictive checking with real-time vision-based reactive avoidance and replanning, specifically designed for dual-arm manipulation.
Closed-Loop TAMP Coupling: The system does not treat task and motion planning as separate stages. It continuously updates the task sequence and allocation based on real-time scene feedback, ensuring robustness against perception errors and dynamic changes.
Real-World Validation in EV Disassembly: The framework is validated on a physical dual-robot setup dismantling Nissan e-NV200 battery packs, demonstrating scalability and safety in a complex, unstructured environment.

4. Experimental Results

The framework was evaluated through three case studies and compared against a baseline Default-RRTConnect planner under identical perception and task assignments.

Trajectory Efficiency:
- Path Length: The proposed system reduced the cumulative end-effector path length from 48.8 m to 17.9 m (a 63.3% reduction). This indicates significantly more compact and efficient motions.
- Makespan: Total execution time improved from 467.9 s to 429.8 s (an 8.1% reduction).
Safety and Workspace Utilization:
- Swept Volume: The volume of space traversed by the robots was drastically reduced.
  - Robot 1: $0.583 \to 0.139 \, m^3$ (76% reduction).
  - Robot 2: $0.696 \to 0.252 \, m^3$ (64% reduction).
- Overlap: The mutual overlap volume between the two robots decreased from $0.064 \to 0.034 \, m^3$ (47% reduction), indicating better spatial coordination.
Statistical Significance: The reduction in path length was statistically significant ( $p < 0.001$ ). The GMM-informed planner also demonstrated lower variance in path deviations and better manipulability (avoiding singular configurations) compared to the baseline.

5. Significance and Conclusion

This work demonstrates that integrating learning-based motion priors with reactive safety layers creates a robust system for multi-robot disassembly.

Practical Impact: The system successfully handles the complexities of EV battery recycling, a critical step for the circular economy, by ensuring safe, efficient, and adaptable automation.
Technical Advancement: It proves that "closed-loop" TAMP, where perception continuously drives re-planning and task re-allocation, is superior to open-loop pipelines in dynamic, shared workspaces.
Future Directions: While effective, the authors note limitations in handling perception outages (occlusions) and potential deadlocks. Future work aims to improve failure recovery mechanisms and explore jointly learned dual-arm motion priors.

In summary, the paper presents a scalable, safety-critical framework that significantly improves motion compactness and coordination efficiency for complex multi-robot tasks, validated through rigorous real-world experimentation.