Research map / Point clouds / Point cloud classification
BiAssemble: Learning Collaborative Affordance for Bimanual Geometric Assembly
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BiAssemble, a bimanual point-level collaborative affordance framework that predicts disassembly directions, alignment transformations, and grasp points for long-horizon geometric assembly of broken parts, and introduces a real-world scanned benchmark.
Problem
Robots must reassemble broken fragments (geometric assembly) using two arms, which requires grasp points that avoid seams, collision-free alignment, and coordinated long-horizon bimanual actions, with evaluation hard to standardize across varied real fragments.
Approach
Extends point-level bimanual affordance (DualAfford) with awareness of subsequent alignment and assembly steps: a disassembly predictor (SO(3)-equivariant VN-DGCNN + cVAE) and a transformation predictor produce alignment poses, and the affordance/critic is trained using whether the downstream controller can reach those poses and assemble without collisions.
Key results
- Ours achieves 24.10% average success on novel instances in training categories versus 21.20% for the w/o SE(3) ablation, 8.40% for DualAfford, 4.20% for Heuristic, and 0.30% for ACT.
- On unseen categories, Ours reaches 17.4% average success versus 14.4% for w/o SE(3) and 7.2% for DualAfford.
- Real-world experiments on two Franka Panda arms with real scanned fragments show qualitative success, and the method extends to three-fragment assembly at 16.6% average success.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the full BiAssemble framework instead of DualAfford | Bimanual geometric assembly · Novel instances in training categories | success rate: 8.40% -> 24.10% | Helped |
| use the full BiAssemble framework instead of DualAfford | Bimanual geometric assembly · Unseen categories | success rate: 7.2% -> 17.4% | Helped |
| add the SO(3)-equivariant VN-DGCNN encoder instead of PointNet++ (ablation) | BiAssemble · Novel instances in training categories | success rate: 21.20% -> 24.10% | Helped |
| add the SO(3)-equivariant VN-DGCNN encoder instead of PointNet++ (ablation) | BiAssemble · Unseen categories | success rate: 14.4% -> 17.4% | Helped |
| provide ground-truth disassembly direction and transformation instead of predicted targets (ablation) | BiAssemble · Novel instances in training categories | success rate: 22.40% -> 24.10% | Hurt |
| provide ground-truth disassembly direction and transformation instead of predicted targets (ablation) | BiAssemble · Unseen categories | success rate: 17.4% -> 19% | Helped |
| use the Affordance Network instead of randomly sampling a contact point (ablation) | BiAssemble · Novel instances in training categories | success rate: 4.60% -> 24.10% | Helped |
| use the Affordance Network instead of randomly sampling a contact point (ablation) | BiAssemble · Unseen categories | success rate: 2.80% -> 17.4% | Helped |
| include the Transformation Predictor (ablation) | BiAssemble · Novel instances in training categories | success rate: 7.40% -> 24.10% | Helped |
| include the Transformation Predictor (ablation) | BiAssemble · Unseen categories | success rate: 4.80% -> 17.4% | Helped |
| use the learned Disassembly Predictor instead of the relative direction between part centers (ablation) | BiAssemble · Novel instances in training categories | success rate: 19.70% -> 24.10% | Helped |
| use the learned Disassembly Predictor instead of the relative direction between part centers (ablation) | BiAssemble · Unseen categories | success rate: 15.20% -> 17.4% | Helped |
Surprising findings
- Providing ground-truth disassembly direction and transformation (w/ GT Target) yields better performance on unseen categories (19%) than the learned method (17.4%).
- Imperfect imaginary assembled shapes from a pretrained pose estimator yield nearly comparable success (20.80% vs 24.10% on training categories).
Limitations
- Absolute success rates remain low (about 17-24%), so the task is far from solved.
- Assumes a perfect imaginary assembled shape as input, with only limited evaluation of imperfect shapes.
- Grasping heavy, smooth, or seam-region fragments still fails; failure cases include grasping seams causing assembly collisions.
Experimental setup
Datasets: Breaking Bad Dataset (EverydayColorPieces)
Benchmarks: BiAssemble simulation benchmark (Breaking Bad EverydayColorPieces, novel instances and unseen categories) real-world scanned fragment benchmark
Metrics: assembly success rate (distance and rotation thresholds)
Baselines: ACT Heuristic SE(3)-Equiv DualAfford w/o SE(3) w/ GT Target
Builds on: DualAfford PointNet++ VN-DGCNN Breaking Bad Dataset
Closest papers by meaning
- GREAT: Geometry-Intention Collaborative Inference for Open-Vocabulary 3D Object Affordance Grounding (CVPR 2025 · similarity 0.88)
- Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation (CVPR 2025 · similarity 0.87)
- AffordDP: Generalizable Diffusion Policy with Transferable Affordance (CVPR 2025 · similarity 0.87)
- Probing and Bridging Geometry-Interaction Cues for Affordance Reasoning in Vision Foundation Models (CVPR 2026 · similarity 0.87)
- SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model (CVPR 2025 · similarity 0.87)
- IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.