FurnitureVLA: Furniture-Centric VLA Systems

• FurnitureVLA is a furniture-centric vision-language-action system that fuses multimodal inputs for real-scale, long-horizon robot assembly and layout generation.
• It employs a progress-aware control mechanism with continuous progress signals and temporal ensembling to enable smooth subtask transitions under challenging geometric tolerances.
• The framework demonstrates robust simulation-to-real transfer and benchmarking across embodied assembly, diagram-video alignment, and symbolic furnishing systems.

FurnitureVLA denotes furniture-centric vision-language-action research in which perception, instruction grounding, and executable decision-making are organized around furniture assembly, furnishing, or layout generation. In its most explicit current usage, it refers to a progress-enhanced Vision-Language-Action policy for real-scale bimanual furniture assembly that jointly predicts robot actions and a continuous progress signal for automatic subtask transitions (Ma et al., 1 Jul 2026). The surrounding literature supplies three closely related strata: embodied assembly environments such as the IKEA Furniture Assembly Environment, fine-grained evaluation suites for assembly understanding, and symbolic or geometric action systems for furnishing and decoration (Lee et al., 2019). This suggests that the term is not fully standardized: it names both a specific long-horizon assembly model and a broader furniture-specific multimodal systems agenda.

1. Terminology, scope, and problem setting

In the assembly literature, FurnitureVLA is defined most concretely as a VLA system for real-scale, long-horizon, bimanual furniture assembly. The task spans three IKEA furniture types—LACK side table, KALLAX shelf, and IVAR chair—with up to 7 semantically grounded subtasks and trajectories as long as 1550 control steps at 10 Hz. The central difficulty is not merely manipulation, but the combination of tight geometric tolerances, persistent dual-arm coordination, frequent occlusion, and compounding error across long horizons (Ma et al., 1 Jul 2026).

A broader interpretation appears across adjacent work on assembly understanding, mixed-reality assistance, furnishing, and layout synthesis. Flat-Pack Bench is explicitly framed as an evaluation bed for furniture-focused vision-language agents that must order multi-step actions, localize assembly events, understand part mating, and track visually similar parts over long videos (Chetan et al., 20 May 2026). IKEA-Bench studies the cross-depiction alignment problem between wordless IKEA diagrams and real assembly video, which is directly relevant to progress monitoring and step-by-step guidance in mixed reality (Liu et al., 1 Apr 2026). In layout generation, the same label has also been used for a structured graph variational autoencoder for indoor furniture layout generation, indicating a second usage centered on structured scene synthesis rather than robot control (Chattopadhyay et al., 2022).

The common denominator across these usages is furniture-specific multimodal grounding. Vision may take the form of egocentric RGB views, top-down floor plans, or diagram images; language may encode subtask instructions, reasoning traces, or design intent; action may be low-level robot control, symbolic placement programs, or constrained geometric arrangements. The literature therefore treats FurnitureVLA less as a single architecture than as a class of systems in which furniture provides the compositional object domain and action semantics.

2. Progress-enhanced VLA for long-horizon bimanual assembly

The 2026 "FurnitureVLA" model is built on the $\pi0.5$ VLA backbone with full finetuning, a PaliGemma–SigLIP vision backbone, and a PaliGemma/Gemma language stack (Ma et al., 1 Jul 2026). Its inputs are four RGB views—front, rear, and two wrist cameras—upsampled from $224 \times 224$ to $448 \times 448$, together with concise semantically grounded subtask instructions and optional inlined robot-state tokens. No observation history is used. The action space is bimanual absolute end-effector control: $a_t^L, a_t^R \in \mathbb{R}^7$, concatenated into $a_t \in \mathbb{R}^{14}$, where each arm uses $[x, y, z, u, v, w, \gamma]$ for pose target and gripper state.

The defining innovation is a scalar progress head $p_t \in [0,1]$ appended to the action target, producing augmented targets of dimension 15. The model decodes chunks of $H=50$ future actions with flow matching and jointly predicts per-timestep progress. Continuous progress is not assigned per subtask as a constant label; it is interpolated over action primitives within a subtask:

$$p_t = \frac{i}{N_k} + \frac{1}{N_k} \cdot \frac{t-s_i}{s_{i+1}-s_i}, \quad s_i \le t < s_{i+1}.$$

A subtask transition is triggered when $\hat p_t \ge \tau_p$, with $224 \times 224$0 and hysteresis filtering. This design is paired with semantically grounded subtask finetuning in which task boundaries are placed at post-retreat, contact-free states, narrowing the initial-state distribution of the next stage and reducing cross-subtask covariate shift.

Inference combines chunk decoding with temporal ensembling. The system executes only the first 5–25 steps of each predicted chunk before replanning, and overlapping chunks are ensembled with exponential weights. The reported effect is a practical coupling of smooth low-level control and responsive stage detection. Relative pose evaluation is carried out against the base part using $224 \times 224$1, with success defined by furniture-scale translation and rotation thresholds. The assembly setting therefore differs from many VLA benchmarks that emphasize short-horizon pick-and-place: the central problem is sustained precision under long temporal dependence, not isolated action selection.

3. Simulation substrate, task formalization, and data generation

The most influential simulation substrate for furniture-centric manipulation is the IKEA Furniture Assembly Environment (IKEA-FAE), introduced as a standardized simulator for long-horizon, hierarchically structured manipulation (Lee et al., 2019). IKEA-FAE supports more than 80 IKEA-inspired furniture models, uses MuJoCo for physics and Unity3D for rendering via DoorGym, and provides three agents: Cursor, Sawyer, and Baxter. Tasks are inherently compositional: an agent repeatedly selects two compatible parts, grasps them, aligns their connectors, and performs attachment through a high-level connect action. Observations include RGB, depth, segmentation masks, a goal image, proprioception, configurable part poses, and scene graph labels.

The environment formalizes attachability through connector alignment checks. For connectors $224 \times 224$2 and $224 \times 224$3 on different parts, attachment depends on positional proximity and orientation agreement:

$224 \times 224$4

$224 \times 224$5

and

$224 \times 224$6

If Attachable = 1 and the agent triggers connect, MuJoCo welds the parts. This gives FurnitureVLA-style systems explicit relational structure: connector IDs, pose thresholds, and scene graphs can be paired with language templates or parsed manuals to create instruction-conditioned policies.

The 2026 assembly system departs from IKEA-FAE in scale and embodiment. Its simulation pipeline is built in Isaac Gym, uses furniture assets from 3D Warehouse, textures from ambientCG, and expert demonstrations generated by motion planning for single-arm actions and object-centric planning for bimanual actions (Ma et al., 1 Jul 2026). Magnet-based attachment is modeled by pose-reset once alignment tolerances are met. The controller runs at 10 Hz; data generation uses stricter thresholds than evaluation, and the sim pipeline is complemented by a VR teleoperation system built around Meta Quest 3 controllers and dual Kinova Gen3 arms. Demonstrations are recorded at 10 Hz, filtered in DROID style, and enriched by mirrored dual-arm modes, decoupled translation and rotation, clutching, and predefined grasp orientation presets. In practice, this creates a two-tier substrate for FurnitureVLA research: synthetic scale and evaluation in simulation, then real demonstrations and rollouts for transfer.

4. Empirical performance and design factors

FurnitureVLA is evaluated against two direct baselines: zero-shot $224 \times 224$7, which fails with 0% success, and a monolithic finetuned $224 \times 224$8 with a single global instruction and no progress modeling (Ma et al., 1 Jul 2026). The monolithic finetune achieves average simulation success of approximately 48%, with LACK at 0.91, KALLAX at 0.11, and IVAR at 0.41. FurnitureVLA raises the average to approximately 80%, with LACK at 0.98, KALLAX at 0.85, and IVAR at 0.56. An additional design-factor study contributes roughly 21% more average success through temporal ensembling, action-horizon selection, rear-view usage, and higher input resolution.

The ablations make two points especially clear. First, demonstration scale matters: 25% of demonstrations yields average success 0.50, 50% yields 0.68, and 100% yields 0.80. Second, the progress signal must be continuous. A discrete progress label per subtask fails completely, with 0% success, because near-completion and completed states are visually similar enough to make discrete stage detection unreliable. Continuous primitive-based interpolation is therefore not a cosmetic auxiliary target; it is the mechanism that enables automatic stage transitions without an external estimator.

Real-world validation is performed on IVAR with a Kinova Gen3 platform. Full-assembly success across subtasks decreases monotonically from 0.80 to 0.40 over the seven stages, while per-part success evaluated independently is 0.80, 0.80, 0.73, 0.80, 0.67, 0.87, and 0.80 (Ma et al., 1 Jul 2026). The paper reports only an approximately 16 percentage-point drop on the hardest IVAR task relative to simulation. The reported gap is attributed to higher noise, occlusions, and multi-point alignment precision, but the results still indicate that progress-aware subtasking, rear-view sensing, and chunked action decoding close much of the sim-to-real gap for real-scale furniture assembly.

5. Benchmarking perception, temporal reasoning, and instruction alignment

Flat-Pack Bench isolates the spatio-temporal competencies that a furniture-focused VLA must possess in real assembly videos (Chetan et al., 20 May 2026). Built on IKEA-Manuals-at-Work, it contains 50 furniture assembly videos covering 24 unique furniture items and 602 multiple-choice questions distributed across Temporal Localization, Temporal Ordering, Mate, and Track. Questions depend on 1–2 visually prompted frames with segmentations and numeric labels, and often require long temporal context: on average, 113.7 frames are needed to answer a question, while Track intervals average 141.1 frames. Human performance reaches 94.18% micro-average, with per-task scores of 93.54% on TOrd, 93.20% on TLoc, 93.77% on Track, and 97.70% on Mate. Current LVLMs remain far below this level: GPT-5 attains 37.71% micro, Gemini 2.5 Pro 33.72%, Gemini 3.1 Pro 32.89%, InternVL3-78B 41.03%, and Qwen2.5-VL-72B 40.37%. The dominant failures are weak temporal utilization, limited tracking, and poor contact reasoning. Image-only evaluation causes large Track drops but can leave Mate nearly unchanged, indicating shortcut reliance on image cues and commonsense rather than video-time reasoning. Zero-shot CoT and Self-Consistency CoT further degrade performance.

IKEA-Bench studies a different but complementary bottleneck: cross-depiction alignment between abstract IKEA diagrams and photorealistic assembly video (Liu et al., 1 Apr 2026). The benchmark contains 1,623 questions across 6 task types on 29 IKEA products, built from 97 assembly videos and 2,569 uniformly sampled frames. Nineteen VLMs are evaluated under Visual, Visual+Text, and Text Only strategies. The main findings are that instruction understanding is recoverable via text, but text simultaneously degrades diagram-to-video alignment; architecture family predicts alignment accuracy more strongly than parameter count; and video understanding remains the hard bottleneck. On Step Recognition in the Visual setting, open-source models range from 33.4% to 59.4%, while Gemini 3 Flash reaches 65.3%. The average drop from T1 Step Recognition to T4 Next-step Prediction is 12.2 points for open-source models. Strategy effects are sharply asymmetric: on D2 Instruction Comprehension, Visual to Visual+Text yields +19.5 points and Visual to Text Only +23.6, but on T1 the corresponding deltas are −3.1 and −5.1. Mechanistically, diagrams and video occupy disjoint ViT subspaces, with ViT-level CKA near zero in three of the four representative models, and adding text shifts attention away from diagram and video tokens toward text tokens. For FurnitureVLA, the implication is direct: mixed-reality assembly assistance is blocked not only by planning, but by unresolved fine-grained temporal perception and depiction-gap alignment.

6. Broader furniture-centric action systems and symbolic variants

Beyond robot assembly, furniture-centric VLA systems also appear in decoration and furnishing. FurniMAS is a multi-agent system for automatic furniture decoration that takes an empty furniture mesh, a natural-language prompt, and an integer $224 \times 224$9, then produces asset selection, style and material assignment, and precise collision-free arrangement on furniture surfaces (Nguyen et al., 7 Jul 2025). It decomposes the pipeline into an Asset Selector, Stylist, Planner, Validators, Arranger, and Retriever, all coordinated through AutoGen. The Arranger solves a constrained optimization problem over positions and discrete orientations using Gurobi, while assets are retrieved from Objaverse through OpenShape text embeddings. Across 200 furniture items and asset counts of 8, 16, and 32, FurniMAS achieves zero OOB and zero BBL due to hard placement constraints, and it attains the highest GPT-4o scores on Functionality, Layout, Scheme, and Atmosphere relative to adapted baselines. In this setting, action is not robot torque or end-effector pose; it is a validated geometric placement program over surfaces.

Architect-Ant takes an explicitly symbolic route to automatic furnishing of floor plans (Rodionov et al., 9 Jun 2026). It uses a Qwen3.5-9B vision-LLM with per-room LoRA adapters to generate a compact coordinate-based DSL in which each furniture placement is an object token with class and axis-aligned metric box. The model is supervised with procedural reasoning traces that encode door-swing clearance, wall alignment, circulation, fixture compatibility, and room-specific inventory rules, and is refined with Direct Preference Optimization using a deterministic rule-based scorer. On CubiCasa5K OOD evaluation, overall best-of-$448 \times 448$0 score improves from 2.04 in zero-shot to 7.27 after SFT and 7.34 after the full method. The DSL remains editable after generation and can be rasterized for a Flux-based LoRA renderer. Architect-Ant thus instantiates a FurnitureVLA pattern in which vision parses room structure, language supplies explicit reasoning traces, and action is a symbolic program validated by rule execution.

A more generative interpretation appears in the structured graph VAE for indoor furniture layout generation, where FurnitureVLA is defined as a room-conditioned graph-structured variational autoencoder with one latent code per furniture node and an autoregressive prior over those latents (Chattopadhyay et al., 2022). On 3D-FRONT, it reports Category KL 0.01–0.02 across room types and scene classification 0.78–0.88, with average synthesis time 130.26 ms versus 148.51 ms for ATISS. An earlier precursor, "Automatic Generation of Constrained Furniture Layouts," models room composition with a directed graphical model and handles constraints through rejection sampling, giving unconstrained generation at 0.04 s per sample and demonstrating that controllable furniture action can also be cast as conditional stochastic layout synthesis rather than embodiment (Henderson et al., 2017). Taken together, these works broaden FurnitureVLA from robot assembly to executable furniture-centered action representations more generally.

7. Limitations, misconceptions, and open problems

A recurring misconception is that general multimodal competence is already sufficient for robust furniture assistance. The benchmark evidence argues otherwise. On Flat-Pack Bench, state-of-the-art LVLMs remain in the 25–41% micro-accuracy range against 94.18% for humans, with persistent failures in long-range tracking, temporal ordering, and physical contact reasoning (Chetan et al., 20 May 2026). On IKEA-Bench, adding text can improve instruction comprehension while harming diagram-to-video alignment, and architecture family matters more than raw parameter count (Liu et al., 1 Apr 2026). Scaling alone has therefore not solved the fine-grained spatio-temporal and cross-depiction problems that furniture assistants must handle.

The embodied assembly stack also has clear structural limits. IKEA-FAE provides no native language annotations, simplifies attachment through MuJoCo welds, and currently supports only one-to-one connector mappings, so identical parts are not interchangeable in the current release (Lee et al., 2019). The real-scale FurnitureVLA system bypasses screwing through magnets, operates on a fixed-base dual-arm setup, and still fails most often on contact-rich subtasks involving heavy or partially assembled structures (Ma et al., 1 Jul 2026). These are not minor implementation details; they delimit the current scope of what “furniture assembly” means in VLA experiments.

Symbolic furnishing systems introduce a different set of limits. Architect-Ant uses axis-aligned boxes only, and its authors report reward hacking under model-pair DPO, especially in kitchens (Rodionov et al., 9 Jun 2026). FurniMAS currently focuses on on-surface placement, with more complex operations such as hanging and draping left open (Nguyen et al., 7 Jul 2025). The broader picture is that FurnitureVLA now spans several mature subproblems—assembly control, temporal video understanding, diagram-video alignment, symbolic furnishing, and structured layout generation—but no single system yet unifies all of them. The literature suggests that the next major advances will require better temporal encoders, stronger part-level tracking, richer geometric and contact reasoning, tighter language grounding, and more faithful physical interaction models.