From CAD to POMDP: Probabilistic Planning for Robotic Disassembly of End-of-Life Products

TL;DR

The paper tackles robotic disassembly under end-of-life uncertainty by formulating disassembly as a POMDP that captures hidden EOL conditions affecting how parts can be separated.It automatically grounds the POMDP from CAD data, robot capabilities, and inspection results, and solves it with a reinforcement-learning policy supplemented by a Bayesian belief updater during execution.Across three products and two robotic setups, the probabilistic planner outperforms deterministic baselines in average disassembly time and variance, and adapts to deviations such as missing or stuck screws.This approach enables more robust, scalable disassembly for circular economy applications and highlights avenues for richer, belief-dependent planning in future work.

Abstract

To support the circular economy, robotic systems must not only assemble new products but also disassemble end-of-life (EOL) ones for reuse, recycling, or safe disposal. Existing approaches to disassembly sequence planning often assume deterministic and fully observable product models, yet real EOL products frequently deviate from their initial designs due to wear, corrosion, or undocumented repairs. We argue that disassembly should therefore be formulated as a Partially Observable Markov Decision Process (POMDP), which naturally captures uncertainty about the product's internal state. We present a mathematical formulation of disassembly as a POMDP, in which hidden variables represent uncertain structural or physical properties. Building on this formulation, we propose a task and motion planning framework that automatically derives specific POMDP models from CAD data, robot capabilities, and inspection results. To obtain tractable policies, we approximate this formulation with a reinforcement-learning approach that operates on stochastic action outcomes informed by inspection priors, while a Bayesian filter continuously maintains beliefs over latent EOL conditions during execution. Using three products on two robotic systems, we demonstrate that this probabilistic planning framework outperforms deterministic baselines in terms of average disassembly time and variance, generalizes across different robot setups, and successfully adapts to deviations from the CAD model, such as missing or stuck parts.

Figures (9)

• Figure 1: Products used in the validation of the system. (a) angle grinder, (b) electric motor with two lids, and (c) electric motor with a single lid.
• Figure 2: Block diagram of the proposed system colored according to sensing (blue), planning (yellow), and execution (red) components. Disassembly is decomposed into a symbolic task planner based on a POMDP formulation and a motion planner that generates executable trajectories.
• Figure 3: Surface plot of the disassembly relation $\psi_{12}(\theta,\phi)$ between the head (1) and the body (2) of an angle grinder. $(\theta,\phi)$ describe the movement direction in spherical coordinates $(\theta,\phi)$. The head can only be pulled in one direction (modelled as $\psi_{12}(\theta,\phi)=1$) resulting in a single green area.
• Figure 4: Illustration of how transition function $T$ changes $\Psi$ (individual $\psi_{ij}(\theta,\phi)$ shown as surface plots) using a rivet as example. Initially, $\prod_{j=1}^{n} \psi_{ij}(\theta,\phi)$ is zero (red in each surface plot) for all directions $(\theta,\phi)$, meaning neither the rivet nor the top can be removed non-destructively. After applying the milling action $a_{mill}$, the relations $\Psi_{Rivet,Top}$ and $\Psi_{Top,Rivet}$ now feature new directions where $\prod_{j=1}^{n} \psi_{ij}(\theta,\phi)=1$ (green), enabling removal of either part with a manipulation action $a_{manip}$. Disassembled parts are represented by $\psi_{ij}=1~\forall j$.
• Figure 5: The two robotic systems used for the experiments with their respective capabilities.