eGRAP: Graph-Based Adaptive Planning

• Electronic-device Graph-based Adaptive Planning (eGRAP) is an integrated, vision-driven framework that uses directed graphs to coordinate dual-arm robotic disassembly for e-waste.
• It leverages real-time RGB-D perception, online graph updates, and a dual-arm scheduling algorithm to adapt to part variations and optimize parallel actions.
• The framework achieves robust, device-agnostic disassembly with high success rates while addressing challenges such as perception noise and hardware limitations.

Electronic-device Graph-based Adaptive Planning (eGRAP) is an integrated framework for autonomous dual-arm robotic disassembly of electronic devices, characterized by a live-updated, vision-driven, and precedence/access–encoded directed graph. Designed to address the challenges of flexible, high-reliability e-waste disassembly, eGRAP systematically integrates perception, planning, and coordinated action execution. The approach targets device-agnostic deployment, adaptability to part and assembly variations, and scalability to new product classes, with specific emphasis on the coordinated management of tooling and manipulation arms through online topological graph sequencing and dynamic task allocation (Das et al., 21 Jan 2026).

1. Problem Context and System Goals

Electronic waste presents one of the fastest-growing global waste challenges, with recycling rates remaining disproportionately low. Controlled, part-aware robotic disassembly is necessary for maximal material recovery but is hindered by the lack of flexible automation. Previous industrial systems rely on fixed scripts, product uniformity, and are brittle to the variability endemic in post-consumer devices. eGRAP addresses these challenges by enabling:

• Device-agnostic operation: Adaptation to new products requires only updates to part labels, rule sets, and perception models.
• Perception-driven adaptability: Real-time graph construction and updating from live RGB-D detection, enabling responsiveness to missing, misaligned, or newly revealed parts.
• Dual-arm coordination: Explicit modeling and scheduling of two arms—one for tooling (screwdriver, with eye-in-hand vision) and one for manipulation (gripper or holder)—maximizing parallelism and honoring hold–operate constraints.

2. Directed Graph Representation of Disassembly State

eGRAP models the current device state as a directed graph $G = (V, E)$, in which:

• $V$: Each node represents a detected part instance (e.g., screw, lid, PCB).
• $E \subseteq V \times V$: Directed edges encode constraints:
  • Precedence: $u \rightarrow v$ denotes that part $u$ must be removed before $v$.
  • Access: $u \rightarrow v$ may also encode physical obstruction; $u$ must be removed to access $v$.

Graph construction is rule-driven at the class level. For instance, in a hard drive, all fastener-class screws are made predecessors of the lid node: $(f \rightarrow L) \in E$ for all $f$ of class Fastener and the unique $L$ (lid). Upon confirmed removal of a node $v$, the system deletes $v$ and its incident edges, maintaining $G$ as a faithful model of the current assembly state.

3. Topological Sequencing and Task Readiness

Disassembly proceeds according to partial orders defined by $G$. At each cycle, the “ready set” $R$ comprises nodes with no unresolved predecessors: $$\mathrm{indegree}(v) = \left|\{ u \in V : (u \rightarrow v) \in E \}\right|,\quad R = \{ v \in V : \mathrm{indegree}(v) = 0 \}$$ Actions corresponding to nodes in $R$ can be carried out without violating precedence or access. Where multiple nodes are available, eGRAP applies class-priority tie-breakers (e.g., Fastener $\succ$ Lid $\succ$ Internals) and estimated move cost minimization, grouping proximate actions. This incremental batch scheduling permits robust, opportunistic parallelism.

4. Dual-Arm Scheduling and Execution Algorithm

eGRAP maintains separate pools of ready actions for:

• Tooling arm (Screwdriving): Primarily “unscrew” actions.
• Manipulation arm (Gripping/Removing): Actions such as “lift,” “remove,” “drop.”

Each iteration of the scheduler carries out:

1. Collection of all ready actions $A = \{a_1, ..., a_k\}$.
2. Capability-based partition into $A_T$ (tooling) and $A_M$ (manipulation).
3. Construction of a compatibility graph over $A_T \cup A_M$ where actions are in conflict if their corresponding nodes are spatially overlapping or share a precedence relationship.
4. Maximal non-conflicting matching: At most one action per arm is assigned, maximizing parallelism.
5. Dispatch of matched actions; unmatched actions remain in the queue.

The above control loop is expressed via the following (simplified) pseudocode: $\begin{aligned} \textbf{while } V \neq \emptyset: & \ \ \text{Perceive and update } G(V,E). \ & \ R \leftarrow \{ v \in V : \mathrm{indegree}(v) = 0 \}. \ & \ A \leftarrow \{ \textsc{InstantiateAction}(v) \mid v \in R \}. \ & \ A_T, A_M \leftarrow \text{split by capability}(A). \ & \ \text{Select non-conflicting pairs from } A_T \cup A_M. \ \textbf{for each assigned action } a: & \ \text{send to the corresponding arm}. \ & \ \text{Wait for completion. On success: remove node from } V. \ & \ \text{On unscrew failure: re-enqueue with small pose offset.} \end{aligned}$ Action primitives are instantiated with predefined approaches, tool parameters, and arm tags.

5. Vision System and Perceptual Feedback

eGRAP employs a two-stage, eye-in-hand vision pipeline based on YOLOv11, trained on approximately 250 manually annotated images.

• Stage 1 (Coarse localization):
  • Intel RealSense D435i provides RGB and depth data.
  • YOLOv11 detects part classes and outputs bounding box, confidence, and 2D centroid.
  • Depth back-projection and hand–eye calibration yield a 3D world-frame approach pose.
• Stage 2 (Fine alignment for screw heads):
  • Micro-camera (aligned to the tool’s axis) images the screw recess.
  • YOLOv11 re-detects the screw head; pixel offset from the center is used for in-plane correction.
  • Guarded motion, with back-off and small pose offset retry on mis-seating.

Larger parts (e.g., lid, PCB, case) employ only Stage 1 for pose estimation. All vision updates integrate tightly with $G$, allowing real-time graph corrections upon new part detection or confirmed disappearance.

6. Online Graph Update and Replanning

The directed graph $G$ is revised online:

• New part detections augment $V$; removed parts are pruned upon action feedback confirmation.
• All edges $E$ are recomputed according to class-level rules over the current $V$.
• This ensures that the ready set $R$ and the action plan remain correct without requiring a reset or global recomputation on each cycle.

This suggests robust operation even under significant perception noise or unmodeled part variability, as the planning policy adapts immediately to the visible state.

7. Experimental Evaluation and System Analysis

Testbed Configuration:

• Tooling arm: 5-DoF xArm5 with custom electric screwdriver (Torx T8), eye-in-hand RealSense D435i, and 0.3 MP micro-camera.
• Manipulation arm: 6-DoF UF850 with vacuum gripper.
• Workpieces: 3.5" hard drives from Samsung, Seagate, Western Digital (7–24 fasteners, layered construction).

Performance Metrics:

	Precision	Recall	[email protected]	Mean 2D Error (px)
Vision	∼0.9	∼0.87	∼0.9	5.8–7.1

	Layer 1 (min)	Layer 2 (min)	Layer 3 (min)	Total Cycle (min)	Success Rate (%)
WD	6.8	9.2	4.8	20.9	70
Seagate	7.2	9.9	4.9	21.5	90
Samsung	6.9	9.5	4.8	21.9	90
Overall	—	—	—	—	83

Failures originated chiefly from hardware issues (e.g., vacuum seal leaks, mis-seats). Coarse-only vision would clear approximately 43% of lid screws; fine alignment enabled 100% with a moderate time increase.

8. Strengths, Limitations, and Prospective Directions

Strengths:

• Integrated perception–planning–execution loop with provably valid sequences via topological reasoning.
• Dual-arm parallelism without reliance on custom disassembly scripts.
• Device-agnostic disassembly model requiring only minimal adaptation for new products.
• Dynamic updates support robust response to missing or unexpectedly revealed parts.

Limitations:

• Limited to rigid parts and standard fasteners; flexible elements (e.g., cables, gaskets) are not currently supported.
• Sensitivity to surface reflections may hamper vision in cases of pronounced glare.
• Greedy action tie-breakers lack a formal global optimality guarantee.

Potential Extensions:

• Incorporating uncertainty-aware edge weights and replanning heuristics for improved robustness to execution failures.
• Extending the graph representation to hierarchical modules, supporting code reuse across product families.
• Modeling compliant or deformable components by augmenting node and edge semantics, along with perception upgrades.
• Deploying eGRAP within larger, cell-level workflows, featuring job-level scheduling and safety integration.
• Benchmarking against standardized e-waste tasks for throughput and reliability comparison (Das et al., 21 Jan 2026).

An implication is that eGRAP's framework generalizes to modular, perception-driven disassembly in heterogeneous manufacturing contexts, contingent upon extensions to handle compliant and variable parts.

---

Reference:

• "Graph-Based Adaptive Planning for Coordinated Dual-Arm Robotic Disassembly of Electronic Devices (eGRAP)" (Das et al., 21 Jan 2026).