FurnitureVLA: Learning Long-Horizon Bimanual Furniture Assembly with Vision-Language-Action Model

At a glance

Problem: errors compound across long assembly sequences. Source
Idea: predict continuous progress alongside actions and transition after retreat. Source
Evidence: simulation on three furniture types and physical IVAR chair trials. Source

Evaluation scope The physical IVAR evaluation uses 100 demonstrations and 15 rollouts, with magnets installed in screw holes. Source

The bimanual data-collection setup includes a teleoperator, cameras and two Kinova arms. Original figure

How it works

Decompose assembly into semantic subtasks and assign continuous progress from action primitives. Source
Filter predicted progress before switching subtasks; clear action buffers when a transition is confirmed. Source

The policy predicts both actions and continuous subtask progress to coordinate transitions. Original figure

Evidence & results

Average full-assembly success in simulation is 80%, versus 48% for monolithic fine-tuning of the same backbone. Evidence
Real IVAR full-assembly success is 40%. Independent per-part success is a different measure and must not substitute for complete assembly. Evidence

Selected successful real-robot rollouts show regrasping and magnet alignment; these examples do not represent the overall success rate. Original figure

Limitations & boundaries

Magnetic connectors simplify the physical assembly setup; these results do not establish autonomous screw fastening. Source
Failures accumulate across subtasks, especially where precise multi-part alignment is required. Source

Takeaway

Explicit progress and carefully chosen transition boundaries help long-horizon control, but complete real-world assembly remains difficult. Source

Resources

Full text ↗PDF ↗

Figures: the paper's authors. Editorial source and image review: 2026-10-21. Findings are author-reported, not independently reproduced by Robotics Research Hub.

Original abstract

Current work on robot furniture assembly mostly focuses on toy-scale settings or single-arm manipulation. We introduce FurnitureVLA, the first systematic study of real-scale bimanual furniture assembly using Vision-Language-Action models (VLAs). We formalize the task, develop a scalable simulation pipeline for expert data generation and evaluation, and build a VR teleoperation system for single-operator bimanual control to collect high-quality real-world demonstrations. To address extreme long-horizon assembly with up to 7 subtasks and 1550 control steps, we propose a progress-enhanced VLA, finetuned on semantically grounded subtasks, that jointly predicts actions and a continuous progress signal, enabling automatic subtask transitions and reducing compounding errors during inference. We further study perception and control design factors that critically affect precision in real-scale assembly. FurnitureVLA improves average simulation success from 48% to 80% compared to baselines across three furniture types, with an additional 21% gain from our design factor study. We validate on a real Kinova Gen3 platform with only 16% drop on the hardest task.

Citation