Anvil: Scaling Embodied Spatial Intelligence through Knowledge Recirculation

Weilin Ruan; Bohan Zhou; Jinghao Li; Yangkai Wei; Wenqian Li; Peiran Dong; Yingjie Xu; Leyi Wu

cs.CVOct 15, 2026

Anvil: Scaling Embodied Spatial Intelligence through Knowledge Recirculation

Authors: Weilin Ruan, Bohan Zhou, Jinghao Li, Yangkai Wei, Wenqian Li, Peiran Dong, Yingjie Xu, Leyi Wu,

Organizations: The Chinese University of Hong Kong · The Hong Kong University of Science and Technology (Guangzhou)

Abstract

Embodied spatial intelligence requires more than visual recognition: models must localize entities, reason about their relations, and express this knowledge in action-relevant forms. Yet joint training can conflate the knowledge needed at different stages because supervision is fragmented across incompatible geometric and semantic outputs. We introduce Anvil, a unified framework for data generation, training, and benchmarking for embodied spatial intelligence. (1) Anvil-Gen is a heterogeneous data flywheel that combines form-specific generation, cross-stage consistency verification, and failure-driven recollection across 11 Ground--Relation--Act forms. (2) Anvil-Ben provides scene-aligned supervision for learning diverse embodied spatial capabilities and scene-linked assessments that distinguish stage competence from error propagation. (3) Anvil-MOPD is a 9B student that consolidates heterogeneous spatial knowledge through distillation from specialist teachers. Semantic-first-fault routing targets teacher supervision to the student's earliest decisive failure, while stage normalization controls the loss scale across native outputs of different lengths. We construct 5.8M source-grounded records spanning 160K images, reserving 580K spatial records for held-out evaluation and using the remaining 5.22M records for training. On average across seven public benchmarks, Anvil-MOPD outperforms Qwen3.5-9B and RynnBrain-8B by 9.29 and 11.03 percentage points, respectively, while reaching 80.67% on VABench-Point. Controlled studies support targeted coverage, selective transfer, and feedback-guided repair as complementary framework components.

Figures & tables

Figure 1: Overview of the Anvil framework. Anvil-Gen produces spatial supervision through generation, cross-verification, and repair. Anvil-Ben curates verified records into training and held-out evaluation sets. Anvil-MOPD consolidates specialist knowledge into a generalist through teacher-and-span routing on student rollouts. The generalist then assists subsequent verification, returning learned knowledge to supervision production while evidence-backed checks retain acceptance authority.

Figure 2: Scene-aligned supervision and training interfaces. The distributions summarize the full pool, including training-only QA replay; the inventory lists eleven native spatial formats and their target signatures. Shared entities connect the left example, while the right gallery illustrates eight formats from different scenes. Appendix Table 5 describes source provenance.

Panel of Table 1 — Table 1: Spatial and embodied-perception evaluation results (%, higher is better).

Figure 3: Controlled tests of coverage, routing, and repair. (a) Coverage matrix at matched updates and record exposures (Context: RoboSpatial; Depth: BLINK). (b) Routing controls at matched maximum rollout budgets (U: uniform; W: weighted). (c) Correct repairs by failure type, using the same

885

candidates and one call each.

Figure 4: Qualitative examples of coverage adaptation and targeted repair. (a) Saved placement predictions before and after adaptation (square: before; circle: adapted; yellow: reference region). (b) Verifier feedback guides a depth-order correction while object boxes remain unchanged. Selected successes illustrate the mechanisms, not their frequency.

Figure 5: Task structure and evidence dependence. (a) Linked-task success before and after closing missing stage tags, with JSON outputs unchanged. (b) Grounding and depth under original, blank, and shuffled images. (c) Plan-validity accuracy on

1{,}012

constructed judgments with

95\%

scene-bootstrap intervals. Panels use the study-specific checkpoints described in Appendix E .

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Family	Form / structure	Native target and consistency requirement
G Ground	$\square$ Box	Referred entity and bounded, ordered corners.
	$\{\square_{i}\}$ Detection	Entity-indexed boxes with non-duplicate instances.
	$\{(x_{i},y_{i})\}$ Mask	Instance or part region consistent with its anchor.
	$\bullet$ Point	Object, contact, or free-space location in a declared frame.
S Relation	$i\prec j$ Depth order	Shared entity set and consistent near–far ordering.
	$d(i,j)$ Metric comparison	Distances or comparisons with supported units.

Appendix

Table 2: Eleven native spatial interfaces. The symbols summarize output structure; colors and family letters identify Ground (G), Relation (S), and Act (A). Each form retains its own validity conditions, while shared entity references connect available forms.

Figure 6: Examples of native spatial supervision. Panels (a–e) and (i–k) show stored annotations from an earlier eight-form pilot. Panels (f–h) illustrate image-plane distance, direction, and graph relations derived from saved entity boxes; panel (l) shows fields rather than an additional task form. Coordinates refer to the original full image on a $0$ – $1000$ scale, including when a panel displays a crop. The examples illustrate recorded supervision and geometry-derived interfaces, rather than model predictions or executed robot trajectories.

Form	Constraint supplied to the generator
Visible point	Preserve the grounded entity label; select visible material rather than a box center by default; return an empty array if no unambiguous point exists.
Mask	Segment the same grounded entity with 8–16 nested coordinate pairs; follow its silhouette rather than returning box corners; abstain when one polygon cannot represent it adequately.
Depth order	Preserve the entity set and labels; rank near to far using visible evidence; image height or apparent size alone is insufficient.
Contact region	Specify a visible functional sub-part and compatible action; place the contact on material rather than an empty container interior.
Gripper	Use the contact proposal, an image-plane jaw axis, an approach class, and a consistent wrist angle; abstain if secure contact is unsupported.

Appendix

Table 3: Representative generation constraints. Concise summaries of the archived templates; the requirements below are not additional experimental measurements.

Target	Evidence to inspect	Reason to withhold or reject
Box / detection	Identity and visible extent of the requested entity.	Wrong instance, omitted material parts, or duplicated detections.
Mask / point	Visible object material and its boundary.	Background inside a polygon; a point in a hole or behind an occluder.
Depth order	Occlusion, contact layering, or a supported perspective cue.	A reversed pair or insufficient evidence to order it.
Affordance / gripper	Named contact part, contact point, and jaw geometry.	A grasp in empty container space or on an unsupported contact surface.
Action plan	Referenced entities, stated preconditions, and ordered steps.	Inapplicable actions or a target inconsistent with the named contact.

Appendix

Table 4: Decision rules used to inspect spatial candidates. The checks concern the visible evidence and the declared task. A valid serialization alone does not establish semantic support.

Source family	Supervision role	Illustrative provenance
Robot video	Grounding, interaction	BridgeData V2 ( Walke et al., 2023 )
Embodied scenes	Spatial targets	RefSpatial ( Zhou et al., 2025 )
Simulation	Objects, actions, physical relations	Embodied-Reasoner ( Zhang et al., 2026d ) ; PhyBlock ( Ma et al., 2025 ) ; MineAnyBuild ( Wei et al., 2025 )
Spatial imagery	Placement, spatial QA, relations	FSD ( Yuan et al., 2026 ) ; SAT ( Ray et al., 2025 ) ; VSI-590K ( Yang et al., 2025c ) ; SpatialSense ( Yang et al., 2019 )
Object states	State-conditioned grounding and actions	ChangeIt frames ( Souček et al., 2022 )
Pointing	Points and caption QA	PixMo ( Deitke et al., 2025 )

Appendix

Table 5: Source families and supervision roles. MIMIC-CGD denotes the CGD subset of MIMIC-IT; VSI-590K is the instruction-tuning corpus released with Cambrian-S.

Study	Updates	Batch	Learning rate	Seeds
Coverage continuation	303	24	$2\times 10^{-5}$	43, 44
Atomic / linked source SFT	512	4	$10^{-4}$	17
Routing continuation	256	4	$5\times 10^{-6}$	17, 23
Matched-size / nested-pool SFT	256	4	$10^{-4}$	17

Appendix

Table 6: Optimization settings for the controlled training studies. Batch denotes the configured effective batch size in records. Coverage exposure counts and realized teacher positions are reported separately.

Model	Any-point (%)	First-point (%)
Base	42.31	36.71
Anvil $\dagger$	81.12	71.33

Appendix

Table 7: Point-cardinality sensitivity on the common 286-item VABench subset with identical image exclusions and saved responses.

\dagger

: earlier source-filtered Anvil checkpoint.

Benchmark	Replay	Vacancy	Mixed
Where2Place	38.00 / 36.22	69.00 / 69.00	68.00 / 66.71
RoboSpatial context	44.14 / 41.74	63.11 / 62.30	62.30 / 59.84
VABench-Point	79.67 / 80.33	77.67 / 76.67	80.67 / 78.33
BLINK depth	90.32 / 90.32	86.29 / 88.71	89.52 / 89.52
ERQA	50.17 / 49.83	46.69 / 46.34	48.08 / 48.08
EmbSpatial	78.43 / 78.24	76.95 / 77.42	77.91 / 77.77

Appendix

Table 8: Source-composition controls (%), seed 43 / seed 44. Updates and row exposures are matched; realized tokens are not. Bold identifies an arm that leads in both seeds.

Comparison	$\Delta$ (pp)	95% CI
Semantic uniform $-$ flat	+2.13	[0.39, 3.90]
Semantic uniform $-$ declared	+0.53	[-0.61, 1.63]
Semantic weighted $-$ uniform	-0.53	[-1.57, 0.42]

Appendix

Table 9: Observed paired routing contrasts on strict Ground–Relation success. Differences are percentage points; 95% scene-bootstrap intervals condition on the two recorded training seeds. Per-seed values and teacher-position counts appear in Table 10 .

Method	G $\to$ S (%)	VQA ratio (%)	Teacher positions
Flat declared-teacher	69.50 / 70.92	100.32 / 99.96	131,334 / 131,704
Structural route	71.28 / 70.92	100.12 / 99.84	221 / 74
Declared route	71.63 / 71.99	100.43 / 99.70	56,267 / 56,449
Semantic (uniform)	72.70 / 71.99	99.70 / 99.97	47,648 / 48,213
Semantic (weighted)	71.99 / 71.63	99.77 / 99.85	48,066 / 48,130

Appendix

Table 10: Per-seed routing observations. Strict G

\to

S success on 282 images / 50 scenes and sampled VQA retention, reported as seed 17 / seed 23. Teacher positions count scored logits including distributed padding; realized budgets are measured, not matched.

Condition	Candidates retained	Fraction of initial pool (%)
Initial candidates	2,011	100.00
Parseable	1,998	99.35
Pass typed checks	1,965	97.71
Source-correct	1,126	55.99

Appendix

Table 11: Observed filtering counts in the shared 2,011-candidate pool. Each percentage uses all initial candidates as its denominator. Source correctness is evaluated against source references, not human precision.

Repair arm	Correct	Yield (%)	GPU-hours	Correct / GPU-hour
Generic retry	11	1.24	2.7664	3.98
Named feedback	111	12.54	2.8186	39.38

Appendix

Table 12: Observed repair productivity. Both arms receive the same 885 rejected candidates and make one call per candidate. GPU-hours measure allocated accelerator time; source-correct recoveries supply the yield numerator.

Initial failure	Candidates	Generic retry	Named feedback
Serialization	13	0 (0.00%)	1 (7.69%)
Typed fields	33	2 (6.06%)	3 (9.09%)
Grounding	761	7 (0.92%)	42 (5.52%)
Depth order	78	2 (2.56%)	65 (83.33%)
Total	885	11 (1.24%)	111 (12.54%)

Appendix

Table 13: Source-correct repairs by initial failure category. The categories partition the rejected pool according to the first failed check; each row reports recovery counts and within-category percentages in both arms.

Pool fraction	Unique records	Input tokens	Success (%)
25%	284	1,524,912	59.93 [49.47, 69.73]
50%	568	1,532,961	58.87 [48.28, 68.77]
100%	1,137	1,535,783	59.22 [50.00, 68.42]

Appendix

Table 14: Nested source coverage at fixed updates. Scene-stratified subsets share fresh-base initialization, 256 updates, 1,024 row exposures, and seed 17. Strict G

\to

S success (%) uses 282 images / 50 scenes; brackets give 95% scene-bootstrap intervals. Tokens are measured.

Figure 7: Failure cases and limits of quality-based learning. (a) A saved placement example regresses after adaptation in both seeds; selected cases do not estimate failure frequency. (b) Parsing, typed validity, and source correctness all use the same 2,011 initial candidates as the denominator. Conditional source correctness among the 1,965 typed-valid records is reported separately in the text. (c) Matched-size feedback training improves the point estimate by 2.84 points, but the paired interval $[-1.41,6.88]$ includes zero.

Shard	Parsed records	Malformed lines	Time (s)
0	167,459	0	3.81
1	163,286	0	3.61
2	167,682	0	3.65
3	100,464	1	2.18
Total	598,891	1	13.26

Appendix

Table 15: Stored-record rechecking: single-run CPU wall time, no model calls.

Training view	Atomic tasks		Linked tasks
	Detection	Depth	Strict G $\to$ S	Tags closed
Ground only	84.75	0.00	18.79	50.71
Relation only	20.60	94.68	0.00	0.71
Atomic mix	84.04	94.33	1.77	44.68
Linked	25.71	0.00	69.86	69.86

Appendix

Table 16: Atomic and linked source training. All four arms use 512 updates and 2,048 row exposures; evaluation uses 282 images from 50 scenes. Detection and Depth use atomic-task prompts, whereas the linked columns use Ground–Relation prompts. Scores are percentages. Tags closed changes only saved closing delimiters; training tokens are recorded rather than matched. Bold marks the highest score in each column.

Model	Benchmark	Original	Blank	Shuffled
Base	Anvil-Bench-G (1,463)	56.87	3.69	3.69
Base	BLINK (124)	90.32	48.39	47.58
Base	RoboSpatial (122)	44.12	2.79	1.80
Anvil $\dagger$	Anvil-Bench-G (1,463)	57.35	7.66	3.49
Anvil $\dagger$	BLINK (124)	89.52	51.61	48.39
Anvil $\dagger$	RoboSpatial (122)	33.60	0.30	4.50

Appendix

Table 17: Visual-input interventions. Full-split scores (%) with questions and targets fixed. Parentheses give sample counts; failures remain in the denominator.

\dagger

: earlier source-filtered Anvil checkpoint, not the final model.

Upstream boxes	Relation specialist	Joint	Specialist $-$ joint
Reference	94.68	94.33	+0.35 [-1.79, +2.43]
Student-produced	93.62	92.20	+1.42 [-1.54, +4.38]
Shifted	84.04	82.62	+1.42 [-2.13, +4.91]

Appendix

Table 18: Upstream localization interventions. Depth accuracy (%) on 282 images from 50 scenes with fixed depth targets. Differences are specialist minus joint, with 95% paired scene-bootstrap intervals (5,000 draws). Both models use 512 source-SFT updates.

Model	Valid	Invalid	Overall $\uparrow$	95% CI
Qwen3.5-9B	40.51	91.11	65.81	[62.07, 70.60]
Anvil $\dagger$	59.09	77.67	68.38	[65.42, 72.57]
Action specialist	55.53	89.53	72.53	[68.22, 78.41]
Joint SFT	53.56	88.54	71.05	[66.86, 76.65]

Appendix

Table 19: State-conditioned plan verification. Accuracy (%) on 1,012 balanced judgments from 506 ALFRED trajectories and 92 scenes. Each valid plan is paired with an invalid ordering of the same actions; 95% intervals resample scenes. State and rules are supplied.

\dagger

: earlier source-filtered Anvil checkpoint, not the final model. Bold marks the highest point estimate in each accuracy column.

Source	Metric $\uparrow$	$N$	Specialist	Joint	RoboBrain
PRISM	Depth agreement	3,601	91.67	88.14	0.00
ALFRED	Exact continuation	3,327	25.91	18.18	0.00
ALFRED	Next-action agreement	3,327	64.44	56.03	0.00
ShareRobot-Bench	Region IoU $\geq 0.5$	193	5.70	6.74	23.32
PRISM	Valid gripper output	580	0.00	0.00	0.00

Appendix

Table 20: Source-Native Preparation and Pretrained Baseline : Percentages, higher is better; Relation specialist for depth, Act otherwise. RoboBrain denotes RoboBrain2.5-8B-NV ( Tan et al., 2026 ) . The ALFRED rows share 3,327 examples. Reference agreement and output validity do not measure execution or grasp success.

Embodied Spatial Reasoning

See all 20 papers

Aug 16, 2026

ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

Spatial intelligence requires foundation models to maintain coherent spatial state across interactions with the physical world. However, existing data-centric approaches typically treat spatial reasoning as independent question-answer instances, enabling shortcut-based answering and providing limited supervision for persistent spatial understanding. To address this, we introduce ChainSpace, a chained-reasoning paradigm that structures spatial reasoning as a state-preserving multi-round process. In this paradigm, spatial questions are organized into logically constrained and jointly consistent chains, where later questions depend on spatial constraints established in earlier rounds. Following this principle, we instantiate ChainSpace-Bench, a manually annotated real-world multi-round benchmark with a Chain-Aware Metric, and ChainSpace-Pipeline, a simulator-based chain-structured supervision generation framework for spatial intelligence training. Experiments show that ChainSpace-Bench exposes chain-level failures that are not captured by isolated question accuracy. Additionally, with a relatively small amount of simulator-generated chained data, models trained by ChainSpace-Pipeline achieve the best performance among open-source models on ChainSpace-Bench and transfer competitively to multiple external spatial intelligence benchmarks. These results establish ChainSpace as an effective paradigm for more faithful evaluation and more data-efficient learning of spatial intelligence.

Xiaohan Zhang, Feng Gu, Xudong Rao +4

College of Information Science and Electronic Engineering, Zhejiang University · Li Auto Inc.

Spatial Reasoning Spatial Reasoning Benchmarks

Ask PDF

Jun 10, 2026

Embodied-BenchClaw: An Autonomous Multi-Agent System for Embodied Spatial Intelligence Benchmark Construction

Benchmarks are essential for evaluating embodied spatial intelligence, yet their construction is labor-intensive, hard to reuse, and difficult to maintain. Existing embodied benchmarks are often static and may quickly become saturated as models improve, limiting their ability to distinguish new capabilities. We propose Embodied-BenchClaw, an autonomous agentic system for constructing embodied spatial intelligence benchmarks. Given a user-specified evaluation intent, Embodied-BenchClaw automatically produces a complete and continually updatable benchmark package through a five-stage pipeline: intent blueprinting, data collection, structuring and cleaning, benchmark synthesis, and evaluation reporting. The pipeline is coordinated by three agents for planning, construction, and evaluation. To improve reusability and reliability, Embodied-BenchClaw introduces an extensible Skill Library and process quality control, enabling benchmark construction to be composable, verifiable, and repairable. We instantiate multiple benchmarks covering indoor spatial reasoning, outdoor spatial reasoning, robotic manipulation, quadruped robot navigation, UAV/aerial-view understanding, and static benchmark enhancement. These benchmarks span diverse embodied carriers, data sources, and spatial capabilities. Experiments with human evaluation, judge-based assessment, consistency checks, cost analysis, and ablations show that Embodied-BenchClaw can construct verifiable, executable, maintainable, and diagnostically useful embodied spatial benchmarks with reduced manual effort.

Baoyang Jiang, Fengchun Zhang, Leyuan Wang +7

QiYuan Lab · School of Information and Software Engineering, University of Electronic Science and Technology of China · Beijing University of Posts and Telecommunications +2

Spatial Reasoning Benchmarks Embodied Spatial Reasoning Synthetic Benchmark Generation Embodied AI Spatial Reasoning

Ask PDF

Sep 30, 2026

Uruqi: Learning Spatial Cognition from Visual Experience

Spatial intelligence requires maintaining a coherent understanding of the world as the embodied agent moves. Like humans, the agent must use its own motion to interpret changes across observations and update object locations and spatial relations accordingly. Despite spatial post-training having substantially broadened the spatial intelligence of vision-language models (VLMs), they still struggle with two atomic spatial capabilities: tracking self-motion and mapping the surrounding world during motion. To address this gap, we provide dense multi-turn supervision over interleaved atomic capabilities within each training episode, mimicking the visual experience of a continuously moving agent that reasons as it observes. To scale this up, we synthesize 11,738 motif-driven camera trajectories over a broad range of 3D scenes, supporting self-motion tracking, persistent object mapping, and rich spatial operations within each visual experience. By training models to reason over these atomic questions, our URUQI

_{\mathrm{Syn}}

-8B improves accuracy from 15.84% to 47.73% on our Uruqi benchmark comprising 52k questions across 2.7k episodes. URUQI-SI-Mix-8B further reaches 50.41%, comparable to the 50.08% achieved by GPT-6 Astra. Trained solely on our synthesized data, URUQI

_{\mathrm{Syn}}

-8B achieves an average relative accuracy improvement of 17.13% over its InternVL3-8B backbone across three external spatial benchmarks. These results highlight continuous visual experience as a scalable source of supervision for developing spatial cognition in VLMs.

Shichao Li, Meiqi Wang, Fei Su +1

Beijing University of Posts and Telecommunications · Tsinghua University

Spatial Reasoning Benchmarks Visual Spatial Reasoning 3D Spatial Reasoning VLM Reasoning Spatial Reasoning Embodied Spatial Reasoning

Ask PDF

Abstract

Embodied Spatial Reasoning

Explore similar work

ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

Embodied-BenchClaw: An Autonomous Multi-Agent System for Embodied Spatial Intelligence Benchmark Construction

Uruqi: Learning Spatial Cognition from Visual Experience