VASR: Vector-Aware Assembly Planning with Equivalent Structures and Repair-Cost Feedback
Organizations: Xi’an Jiaotong-Liverpool University · The Hong Kong University of Science and Technology (Guangzhou)
Abstract
Furniture assembly planning requires inferring subassembly hierarchies from individual parts and a target illustration. Existing raster-based methods do not explicitly exploit the geometric and structural information in SVG drawings. We introduce VASR, a framework that fuses features from a dedicated SVG encoder with visual representations from a pretrained vision-language model to predict assembly hierarchies. To avoid penalizing valid alternatives to a reference assembly tree, we propose Repair-Aware Hierarchical Policy Optimization (RHPO) with hierarchy-preserving equivalence rewards. RHPO further uses repair-cost feedback to favor decisions with lower downstream correction costs among candidates with comparable equivalence-aware rewards. VASR achieves the highest Equivalent-Hard F1 among the evaluated methods, outperforming GPT6-Astra by 17.69 percentage points on the IKEA-derived test set while maintaining strong performance on an independent furniture dataset.
Figures & tables
| Variant | Eq.-Simple | Eq.-Hard | ECS | ECH | ECR |
|---|---|---|---|---|---|
| VASR w/o SVG encoder | 48.82% | 30.31% | 50.14% | 39.67% | 21.72% |
| VASR-S | 50.43% | 33.75% | 56.06% | 48.91% | 24.49% |
| VASR w/o planner components | 38.12% | 19.91% | 36.32% | 24.19% | 11.76% |
| VASR (full) | 67.73% | 54.81% | 77.35% | 62.43% | 33.17% |
| Variant | Repair-cost | Quality | Controlled | Action | Eq.-Hard F1 | ECR |
|---|---|---|---|---|---|---|
| preference | gating | repair | distillation | |||
| Full RHPO | 54.81 0.72% | 33.17 0.85% | ||||
| w/o Repair Cost | 41.30 0.91% | 27.80 1.12% | ||||
| w/o Quality Gating | 48.24 1.05% | 30.45 0.76% | ||||
| w/o Controlled Repair | ⋆ | ⋆ | 44.18 0.87% | 28.62 0.45% | ||
| w/o Action Distillation | 46.72 0.54% | 29.76 1.03% |
| Variant | Success | Quality | Repair Cost | Valid Mass | Invalid Mass | Unknown Mass |
|---|---|---|---|---|---|---|
| Full RHPO | 0.74 0.01 | 0.83 0.01 | 1.20 0.08 | 0.72 0.02 | 0.08 0.01 | 0.20 0.01 |
| w/o Repair Cost | 0.73 0.01 | 0.82 0.02 | 1.78 0.10 | 0.64 0.02 | 0.15 0.02 | 0.21 0.01 |
| w/o Quality Gating | 0.64 0.02 | 0.75 0.02 | 0.95 0.07 | 0.59 0.03 | 0.20 0.02 | 0.21 0.01 |
| w/o Controlled Repair | 0.68 0.02 | 0.78 0.02 | 1.92 0.12 | 0.63 0.02 | 0.16 0.02 | 0.21 0.01 |
| w/o Action Distillation | 0.70 0.02 | 0.80 0.01 | 1.50 0.09 | 0.58 0.02 | 0.22 0.02 | 0.20 0.01 |
Robotic Assembly
Robotic assembly studies how robots perceive, plan, and join parts reliably, from precision insertion to collaborative construction.
See all 93 papers