Organizations: Division of Information Science, Graduate School of Science and Technology, Nara Institute of Science and Technology, 8916-5 Takayama, Ikoma, Nara, 630-0192, Japan
Non-destructive extraction of the target internal part, such as batteries and motors, by cutting surrounding structures is crucial at recycling and disposal sites. However, the diversity of products and the lack of information on disassembly procedures make it challenging to decide where to cut. This study explores a method for non-destructive extraction of a target internal part that iteratively estimates the internal structure from observed cutting surfaces and formulates cutting plans based on the estimation results. A key requirement is to estimate the probability of the target part's presence from partial observations. However, learning conditional generative models for this task is challenging: The high dimensionality of 3D shape representations makes learning difficult, and conventional models (e.g., conditional variational autoencoders) often fail to capture multi-modal predictive uncertainty due to mode collapse, resulting in overconfident predictions. To address these issues, we propose VoxelDiffusionCut, which iteratively estimates the internal structure represented as voxels using a diffusion model and plans cuts for non-destructive extraction of the target internal part based on the estimation results. Voxel representation allows the model to predict only attributes at fixed grid positions, i.e., types of constituent parts, making learning more tractable. The diffusion model completes the voxel representation conditioned on observed cutting surfaces, capturing uncertainty in unobserved regions to avoid erroneous cuts. Experimental results in simulation suggest that the proposed method can estimate internal structures from observed cutting surfaces and enable non-destructive extraction of the target internal part by leveraging the estimated uncertainty.
Figures & tables
Fig. 1 : Overview of the process for non-destructive extraction of the target internal part. The process iteratively estimates the internal structure from the observed cutting surfaces and plans cutting positions to extract the target part without damage.
Fig. 2 : Framework of VoxelDiffusionCut. The internal structure prediction model is trained using a diffusion model based on data with various location patterns of internal parts. During deployment, the model estimates the internal structure through conditional generation based on the observed cutting surfaces and subsequently plans the cutting positions.
Fig. 3 : 3D shape representation by voxelization. (a) Voxel representation with a discrete 3D grid K . (b) Examples of cutting surfaces (slices) extracted along the Y -axis (indexed by j ), with each surface denoted as x(i,j) .
Fig. 4 : Cutting action planning along the Y -axis based on predicted internal structure. Values on each cutting surface indicate the presence score of the target internal part. Positions 2 and 3 are feasible owing to their low presence scores, and 3 is selected because it maximizes the removable volume.
Fig. 5 : Object configurations used for evaluation. The battery is defined as the target internal part for extraction. In the simple-shaped model, Part B corresponds to the battery.
Fig. 6 : Examples of cutting process transitions by the proposed VoxelDiffusionCut with the simple-shaped model (Object A). VoxelDiffusionCut plans cutting positions based on the estimated internal structure to maximize the removable volume while non-destructively extracting the target internal part.
Table 7
Fig. 7 : Rendered objects after processing by each method for the simple-shaped models
Fig. 8 : Examples of cutting process transitions by the proposed VoxelDiffusionCut with the complex-shaped model (Object D)
Table 10
Fig. 9 : Rendered objects after processing by each method for complex-shaped models. The areas enlarged in red frames show where the baseline method erroneously cut the target internal part.
Fig. 10 : Comparison of presence score maps for target internal part locations in complex-shaped models at task step t =8