Venue: ICLR 2026 Category: Benchmark — Spatial reasoning / Embodied VLM evaluation Trend tag: Spatial · 3D for VLA · Multi-view evaluation Affiliations: Tsinghua, PKU, Fudan, Microsoft Research Asia, HKUST, Zhejiang Univ. (Feng, Kang, Wang, Du, Yan, Shi, Yuan, Liang, Deng, Li, Yang, An, Zheng, Wang, Chen, Xu, Liang, Yang, Guo)

flowchart LR
  Src[AgiWorld + BridgeV2<br/>980 episodes, multi-camera] --> Filter[Rule-based + GPT-4.1 filter<br/>candidate triage]
  Filter --> Anno[Human annotation<br/>5-choice MCQ, 8 task templates]
  Anno --> QA[1,708 multi-view QA items]
  QA --> Eval[Zero-shot eval<br/>blind / proprietary / reasoning / open-source / MoE / human]
  Eval --> Cot[CoT-style augmentations<br/>text descriptions, VGGT views, MoGe-2 depth]
  Eval --> Cor[Correlation analysis<br/>spatial vs robotic, single-view vs multi-view]

VLM spatial-reasoning benchmarks are mostly single-view (EmbSpatial-Bench, RoboSpatial, SpatialVLM, VSI-Bench, OmniSpatial, ShareRobot). Real robots use multi-camera setups. The two existing multi-view benchmarks — All-Angles Bench and Ego3D-Bench — focus on photographic alignment or egocentric navigation, not manipulation. ERQA and MMSI-Bench include partial multi-view but limited diversity and basic perception only. The paper's claim: strong single-view scores do not predict multi-view robotic competence, so a dedicated benchmark is needed.

Stage 1 — Data Collection. Source from AgiWorld (single-arm + dual-arm) and BridgeV2 (single-arm). Rule-based filtering for temporal separation, scene diversity, visual clarity. GPT-4.1 triages candidate image pairs against the 8 task definitions. Final human verification.

Stage 2 — QA Generation. Trained annotators write 5-choice MCQs from task-specific templates. GPT-4.1 is never used to generate QA content — only auxiliary triage. Distractors must be plausible but clearly distinguishable. After automatic sampling and manual inspection, more than 3,000 high-quality image pairs were retained as the foundation pool (Appendix F).

Stage 3 — Human-in-the-loop Quality Review. Iterative review checks alignment, sentence structure, image-pair quality, distractor balance, and answer-distribution randomization. Final pool: 1,708 QA items from 980 episodes.

Spatial understanding (4 subtasks):

1. Cross-View Matching — identify the same object across views (e.g., red bbox in right-gripper view → which colored bbox in left-gripper / head view).
2. Distance Judgement — relative distances between objects from multi-view input.
3. Viewpoint Identification — reason about viewpoint transformation (which gripper view matches a given head-camera image).
4. 3D Spatial Consistency — assign coordinate triplets to objects given a head-view-anchored 3D cube.

Robotic execution (4 subtasks): 5. Action Planning — choose a multi-step manipulation sequence in a head-anchored frame. 6. Step Execution — verify whether a specific next-step movement is correct. 7. Trajectory Selection — pick the colored line most likely to complete a task. 8. Affordance Recognition — pick the grasp candidate most likely to succeed.

• Textual CoT (w cot / w text): GPT-4.1-generated scene descriptions appended to input.
• Visual CoT (w vggt): synthesized novel views via VGGT (Wang et al., 2025a).
• Structural CoT (w depth): depth priors via MoGe-2 (Wang et al., 2025b).

• GPT-5 (56.41) leads overall but is 34.6 points below human (91.04).
• Gemini-2.5-pro (49.52) and o4-mini (46.47) are second and third — all reasoning-optimized models.
• Best non-reasoning proprietary: GPT-4.1 (30.90). Best open-source dense: Qwen2.5-vl-72B (24.29). Best open-source MoE: Llama-4-Maverick (26.11).
• 3D Spatial Consistency is the most discriminating subtask: non-reasoning models cluster near random (~6–14%) while reasoning models jump to 49–82%. GPT-5 hits 82.35 here.
• All models trail badly on Affordance Recognition (best non-human is GPT-5 at 39.23 vs human 89.05).
• Single-view ablation (Appendix G, Table 7): restricting to the most informative third-person view (AgiWorld head camera / BridgeV2 view1) on the four degradable subtasks drops GPT-5's Distance Judgement by 18.90 points (multi-view 36.32 here, ∆ +18.90), and GPT-4.1 by +10.44 — confirming the benchmark genuinely tests multi-view fusion. Gains correlate with model capability: smaller models (Qwen2.5-vl-7B/32B) show negligible or negative ∆.

Synthetic novel views (w vggt) tend to hurt performance — the authors note current view-synthesis methods perform poorly under narrow baselines, cluttered tabletops, and gripper-centric viewpoints. Depth priors help only when the backbone has capacity to exploit them (GPT-4.1 yes, Qwen2.5-vl-7B partial). Textual CoT shines for mid-capacity open-source models like Gemma-3-12B but barely moves saturated proprietary models.

• Internal axis (Fig. 5). Spatial vs robotic accuracy: positive correlation, but only proprietary and reasoning-optimized models occupy the upward region. Most open-source models cluster near random on both axes.
• External axis (Fig. 6). OmniSpatial (single-view) accuracy vs MV-RoboBench accuracy: strong single-view performance does not transfer reliably. Many models that score well on OmniSpatial remain near random on multi-view tasks.

The paper does not enumerate "limitations" in a single section but the discussion (Section 6) and appendices flag:

1. Benchmark scale is modest — 1.7k items vs 8.6k for Ego3D-Bench or 5k for VSI-Bench — chosen for manual quality control over template generation.
2. Source-domain bias — derived only from AgiWorld + BridgeV2 (tabletop manipulation). Locomotion, mobile manipulation, and humanoid scenarios are out of scope.
3. Multiple-choice format simplifies grading but cannot probe open-ended generation quality. Future work targets free-form spatial reasoning evaluation.
4. CoT augmentations are off-the-shelf — no fine-tuning of view-synthesis or depth models for robot scenes; current novel-view-synthesis baselines genuinely fail on narrow-baseline gripper views.
5. Looking forward (Sec. 6). Authors call for (i) architectures with explicit geometric priors and cross-view consistency, (ii) training pipelines aligning perception with action grounding, and (iii) larger-scale multi-camera datasets reflecting real-world manipulation complexity.

MV-RoboBench is the first benchmark coupling multi-view perception with robotic manipulation reasoning. It serves as the diagnostic counterpart to method papers like:

• Spatial Forcing — claims to inject spatial competence into VLAs via training.
• SP-VLA — spatially guided VLA inference.
• FALCON — converts spatial reasoning to action chunks.

For each of these, MV-RoboBench provides a way to ask: did the multi-view spatial competence actually transfer, rather than relying on single-image proxies like SpatialVLM or RoboSpatial. It complements (rather than replaces) ERQA and MMSI-Bench by being the first to put manipulation-style multi-view reasoning front and center.

The benchmark's main empirical result — that reasoning-optimized models (o4-mini, GPT-5, Gemini-2.5-pro) win by 15–30 points over non-reasoning peers and that 3D-Spatial-Consistency separates these tiers most cleanly — is a strong signal that multi-view robotic reasoning rewards explicit chain-of-thought, not raw perception scaling. This validates the bet behind systems like RoboBrain, RoboRefer, and Embodied-R1 that pair VLAs with explicit reasoning passes.

• OpenReview: https://openreview.net/forum?id=jXDZJAfRZB
• PDF: https://openreview.net/pdf?id=jXDZJAfRZB
• arXiv: https://arxiv.org/abs/2510.19400
• Code/data: https://github.com/microsoft/MV-RoboBench

• Spatial Forcing
• SP-VLA
• FALCON
• Embodied-R1

← Back to ICLR-2026