We introduce a systematic study of real-scale bimanual furniture assembly using Vision-Language-Action Models (VLAs). We build a scalable simulation pipeline and VR teleoperation system for large-scale data collection, propose a progress-enhanced VLA for improved long-horizon execution, and conduct a focused study of VLA design factors for assembly precision. This work represents a step toward deploying VLAs in challenging, real-world manipulation tasks.
Bimanual Furniture Assembly Tasks
Hardware Setup. Assembly is performed on a tabletop using two Kinova Gen3 7-DoF robot arms, with a Robotiq Hand-E gripper on the left arm and a Robotiq 2F-85 gripper on the right. In addition to the front and wrist-mounted cameras commonly used in prior VLA work, we add a rear camera to provide complementary views, as large furniture parts and frequent bimanual interactions often occlude frontal observations.
Task Properties. We study real-scale bimanual furniture assembly using three IKEA items of increasing difficulty: LACK side table, KALLAX shelf, and IVAR chair. We define subtask decompositions with corresponding language instructions that are concise, descriptive, and semantically grounded. Our tasks exhibit the following challenging properties:
- Long-horizon execution. Each assembly requires completing a strict sequence of subtasks, with horizons ranging from 4 (LACK, KALLAX) to 7 (IVAR), and 650–1550 control steps (65–155 seconds), exceeding prior benchmarks.
- High-precision insertion. Each subtask requires precise alignment between mating parts under tight translation and rotation thresholds, stricter than prior toy-scale assembly benchmarks.
- Diverse manipulation skills. Each subtask is specified by a structured language instruction (e.g., "grasp, align, and insert a part") and requires composing multiple manipulation skills, such as grasping, alignment, insertion, lifting, and rotation. A simple assembly (LACK with 4 subtasks) involves roughly 12 skill executions, while more complex assemblies like IVAR can involve 25 skills due to longer horizons and additional bimanual operations.
- Bimanual coordination. Several subtasks require both arms to act in concert over large, heavy parts. For example, attaching the KALLAX top panel requires simultaneously lifting, rotating, and aligning a full-width panel with both arms. These bimanual subtasks involve up to four sequential action primitives and must maintain reachability and avoid singularities throughout.
Real-World Data Collection via VR Teleoperation
To support real-world deployment, we develop a VR teleoperation system for high-quality demonstration collection, enabling coordinated dual-arm control by a single teleoperator on a Kinova Gen3 platform. Given the complexity of long-horizon, high-precision bimanual assembly, we adopt the following design principles:
- Decoupled translation and rotation control. All three translational axes are controlled simultaneously for smooth end-effector motion, while rotation is handled independently to improve stability and precision. This is critical for precise insertion subtasks, such as aligning rails in IVAR.
- Pre-defined grasp primitives. We define a set of grasp poses with discrete orientation variants, triggered via button inputs. Upon activation, the end-effector orientation snaps to a preset while maintaining its current position, removing manual wrist rotation and enforcing consistent, task-specific configurations.
- Synchronized bimanual control. We introduce a synchronized mode in which both arms execute mirrored commands simultaneously, enabling efficient repositioning, alignment, and rotation of large and heavy furniture components.
Simulation Experiments
We examine: 1) how the progress-enhanced VLA improves long-horizon execution; 2) how key perception and control design factors affect assembly precision and success; and 3) ablations of our design choices.
Implementation Details
Simulation. We use Isaac Gym, extending the FurnitureBench codebase, with furniture 3D models from 3D Warehouse, textures from ambientCG, and assets post-processed in Blender 5.0.
Training and Inference. We use a VLA backbone based on [32], finetuned for 40,000 steps on 8 NVIDIA L40S GPUs with global batch size 64. Inference runs on a single L40S GPU with standard inference parameters.
Long-Horizon Assembly Performance
Setup. We generate 500 demonstrations per furniture for finetuning and evaluate on 100 rollouts each, training a single VLA model across all furniture types.
Evaluation Metrics. We report the success rate of full assembly (i.e., completing all subtasks). If any preceding subtask fails, the rollout terminates, resulting in a monotonically decreasing success rate across subtasks.
Real-World Experiments
We validate our approach on a real bimanual Kinova Gen3 platform, testing on the most challenging IVAR chair.
Implementation Details
We use Quest2ROS to map VR controller poses and inputs to robot end-effector commands. Colored markers (e.g., black and blue tape) are placed on parts to assist the teleoperator with visual alignment during demonstration collection. For the collected teleoperated demonstrations, we filter no-op actions following DROID, but do not trim the initial and trailing frames of each non-idle segment, as they contain critical signals indicating subtask completion.
Per-Part Assembly Performance
Setup and Evaluation. We evaluate each subtask independently. We initialize robot and scene at the corresponding intermediate assembly state and run 15 rollouts per subtask.
Results. Per-part success rates are consistently higher than full assembly rates, confirming that failures accumulate across subtasks rather than stemming from any single catastrophic failure mode. Subtask 5 remains the most challenging because of its high precision demands.
Conclusion and Discussion
We introduce FurnitureVLA, the first systematic study of real-scale bimanual furniture assembly with VLAs. We build a scalable simulation pipeline and VR teleoperation system for large-scale testbed and data collection, propose a progress-enhanced VLA for improved long-horizon execution, and conduct a focused study of VLA design factors for assembly precision. We see this work as a step toward deploying VLAs in challenging, real-world manipulation tasks.
Limitations. Our fixed-base setup limits assembly to furniture within the robot's kinematic workspace; mobile bimanual platforms could handle larger items. We also bypass screwing with magnets, extending to tool use remains an open challenge requiring significantly higher precision.
Appendix Overview
This appendix includes: 1) additional details on constructing the simulation environment, 2) details of the real-world system, and 3) more experimental details on training and results breakdown.
Furniture Models
The IKEA furniture used can be purchased online. 3D models and textures are also publicly available for reproducibility. The furniture items used are: - LACK side table (serial: 30449908) - KALLAX shelf (serial: 90301555) - IVAR chair (serial: 90263902)
Part names are manually defined based on their spatial location relative to the base part in the assembled configuration. In simulation, all parts are assigned a density of pine wood, consistent with real IKEA components.
Additional Experiments and Details
Training Hyperparameters. We perform full finetuning of the VLA backbone using standard hyperparameters.
Simulation Results Breakdown. Subtask success rates from the main simulation tables are visualized in the full results breakdown.
Frequently Asked Questions
How many demonstrations are needed to train FurnitureVLA? We generate 500 demonstrations per furniture type for finetuning, and train a single VLA model across all three furniture types simultaneously.
What makes the IVAR chair the most challenging assembly? The IVAR chair requires 7 subtasks and involves up to 25 skill executions, including bimanual operations like coordinated lifting and rotating the entire partially-assembled chair frame.
How does the VR teleoperation system enable single-operator bimanual control? The system uses decoupled translation/rotation control, pre-defined grasp primitives, and a synchronized mode where both arms execute mirrored commands simultaneously for coordinated manipulation.
What are the main limitations of the current approach? The fixed-base setup limits assembly to furniture within the robot's kinematic workspace, and the system bypasses screwing with magnets, leaving tool use as an open challenge.