Overview
At a glance
Spatial-Interactor trains vision-language models to connect actions and camera motion with changes in what they see. A 107,518-example interaction dataset supports supervised learning of local transitions, followed by on-policy distillation of longer trajectories. Four Qwen backbones improve their four-benchmark average by 16.4–25.0 percentage points. Most gains come from supervised training; the controlled 7B comparison adds 0.9 points over answer-only reinforcement learning. Closed-loop transfer improves but remains limited: WalkerBench success rises from 7% to 14%, and broader embodied diagnostics include regressions.
1 Introduction
Correctly naming objects does not ensure a model can track how a scene changes when an object or observer moves. Existing spatial benchmarks reveal weaknesses in temporal order and viewpoint integration [43] [39] [27] [40]. Shuffling video frames reduces the two tested base models’ scores by less than one point, suggesting limited use of sequence order. Supplying four chronological motion descriptions improves a separate GPT-5.5 displacement diagnostic from 23.9 to 35.5, motivating training that internalizes transition evidence. That diagnostic provides extra text at inference; the proposed student normally receives only the visual input and question.
2 Method
2.1 State-Transition Learning from Spatial Interaction
A visual observation depends jointly on the world state and the observer’s pose. The learning unit is a before/action/after transition; consecutive transitions form a trajectory. Local tasks ask what changed or caused a change, while longer tasks combine these changes into a coherent path. This formulation distinguishes motion in the scene from motion of the camera.
Equation 1
| (1) |
2.2 Progressive Spatial Interaction Curriculum
The curriculum separates three demands. L1 tracks external changes under a stable view, including object operations and their order. L2 tracks the observer in a relatively stable world, including camera movement, cross-view correspondence, parallax, and imagined viewpoints. L3 combines a complete trajectory to infer displacement, distance traveled, turns, revisits, and reverse paths.
Simulator actions and measured camera or robot trajectories provide targets. Deterministic question templates preserve those targets; geometry and answerability checks reject unreliable examples. The frozen visual verifier may discard a question but cannot invent or revise its ground-truth answer.
2.3 Local State-Transition Modeling
Supervised fine-tuning on L1–L2 teaches the relation between an action and its visible spatial consequence. The same transition can support action inference, prediction of a resulting view, or measurement of a change. These short examples supply a foundation for learning from longer sequences.
2.4 Long-Horizon State-Transition Integration with OPD
For long videos, a frozen visual annotator summarizes four consecutive eight-frame intervals without seeing the question or answer. These ordered descriptions are privileged training context. A teacher branch uses a stopped-gradient snapshot of the current policy, refreshed before each update, and receives the descriptions alongside the original input. The student does not receive them.
Both branches score the same student-generated prefixes. A forward KL process loss aligns their distributions over reasoning tokens; final-answer tokens are excluded. Answer rewards train correctness through GRPO, while a separate frozen SFT reference provides the usual KL regularizer. The privileged teacher is a snapshot of the student, not a separately stronger language model. Its guidance gradually decays to zero.
Equation 2
| (2) |
3 Experiments
3.1 Experimental Setup
Training Protocol.
The backbones are Qwen2.5-VL-3B/7B and Qwen3-VL-4B/8B [2] [1]. SFT combines 82,596 L1–L2 training examples with 80,000 external spatial questions. Long-horizon training uses 10,712 RoomTour3D L3 examples and 10,783 VSTI-Bench training examples. Both stages freeze the visual encoder and update the language model and multimodal projector. Scene/video splits are disjoint, but several main benchmark families contribute training data.
Benchmarks.
The main tests are VSI-Bench, MindCube Tiny, VSTI-Bench, and the multi-view portion of SPBench [43] [39] [11] [25]. Videos use 32 ordered frames; multi-view inputs retain all provided views. Official accuracy or MRA scores are reported. “Overall” averages the four benchmark means, so it mixes different task metrics. MMSI-Bench, ViewSpatial-Bench, SAT-Real, and SAT-Syn are reserved for cross-benchmark transfer [45] [24] [35].
3.2 Main Results
The four benchmark averages rise from 33.9 to 55.6 for 3B, 47.3 to 63.7 for 4B, 35.1 to 60.1 for 7B, and 48.4 to 65.9 for 8B. These are gains of 16.4–25.0 percentage points. The 8B edition leads the listed VSI-Bench and MindCube Tiny results, while 7B leads VSTI-Bench; it does not lead every benchmark. Competitor architectures and training mixtures differ, so those rankings do not isolate the method’s contribution.
| Model | VSI | MindCube | VSTI | SPBench-MV | Overall | |||
| Rel. Dist. | Route | Avg. | Avg. | Displ. | Avg. | Avg. | Avg. | |
| Proprietary models | ||||||||
| GPT-5.5 | 55.4 | 39.2 | 45.5 | 63.9 | 23.9 | 48.1 | 56.0 | 53.4 |
| GPT-4o | 37.0 | 31.5 | 34.0 | 36.5 | 23.4 | 38.2 | 48.2 | 39.2 |
| Gemini-2.5-Pro | 61.1 | 45.9 | 51.5 | 52.2 | 8.0 | 42.4 | 58.3 | 51.1 |
| Open-source models | ||||||||
| Qwen2.5-VL-3B | 33.8 | 27.3 | 27.0 | 33.2 | 15.3 | 38.6 | 36.6 | 33.9 |
| Qwen2.5-VL-7B | 36.6 | 30.4 | 37.9 | 29.3 | 8.6 | 39.0 | 34.0 | 35.1 |
| Qwen3-VL-4B | 52.4 | 34.0 | 52.8 | 27.2 | 25.9 | 44.8 | 64.2 | 47.3 |
| Qwen3-VL-8B | 53.5 | 32.0 | 55.7 | 29.4 | 24.2 | 43.3 | 65.1 | 48.4 |
| InternVL3-2B | 32.2 | 32.9 | 32.9 | 37.5 | 21.4 | 32.5 | 37.6 | 35.1 |
| InternVL3-8B | 48.0 | 26.3 | 42.1 | 41.5 | 22.4 | 43.1 | 45.5 | 43.1 |
| Spatial reasoning models | ||||||||
| SpaceR-7B | 39.7 | 31.4 | 44.5 | 29.9 | 12.7 | 43.7 | 56.9 | 43.8 |
| Video-R1-7B | 36.8 | 31.4 | 33.4 | 30.9 | 5.4 | 36.2 | 40.7 | 35.3 |
| VILASR-7B | 45.1 | 29.9 | 44.6 | 35.1 | 15.7 | 46.2 | 57.6 | 45.9 |
| Spatial-MLLM-4B | 40.4 | 33.0 | 46.3 | 32.1 | 16.9 | 38.0 | 58.9 | 43.8 |
| SpatialLadder-3B | 45.6 | 27.3 | 44.8 | 43.5 | 29.7 | 46.2 | 70.9 | 51.4 |
| Spatial-Interactor | ||||||||
| Spatial-Interactor-3B | 53.0 19.2 | 33.0 5.7 | 50.4 23.4 | 59.3 26.1 | 32.4 17.1 | 48.3 9.7 | 64.5 27.9 | 55.6 21.7 |
| Spatial-Interactor-4B | 60.7 8.3 | 36.6 2.6 | 57.2 4.4 | 80.3 53.1 | 31.0 5.1 | 47.1 2.3 | 70.3 6.1 | 63.7 16.4 |
| Spatial-Interactor-7B | 50.6 14.0 | 34.0 3.6 | 49.0 11.1 | 74.3 45.0 | 36.6 28.0 | 49.5 10.5 | 67.6 33.6 | 60.1 25.0 |
| Spatial-Interactor-8B | 62.7 9.2 | 39.7 7.7 | 60.0 4.3 | 88.4 59.0 | 33.2 9.0 | 47.1 3.8 | 68.0 2.9 | 65.9 17.5 |
3.3 Cross-Benchmark Generalization
Across four benchmark families absent from training, the base-to-final mean accuracy gains are 4.3, 6.7, 8.4, and 10.6 points for 3B, 4B, 7B, and 8B. SFT supplies most of this improvement; OPD adds another 1.0–1.4 points. Transfer is therefore broader than the training benchmark families, although individual task gains are uneven.
| Model | MMSI | ViewSpatial | SAT-Real | SAT-Syn | Overall |
|---|---|---|---|---|---|
| Qwen2.5-VL-3B | 29.0 | 35.6 | 56.7 | 56.3 | 44.4 |
| + SFT | 29.2 | 43.2 | 61.0 | 56.8 | 47.6 |
| + SFT + OPD | 29.9 0.9 | 44.3 8.7 | 62.0 5.3 | 58.7 2.4 | 48.7 4.3 |
| Qwen2.5-VL-7B | 27.7 | 36.3 | 54.7 | 57.5 | 44.1 |
| + SFT | 30.4 | 41.3 | 71.3 | 61.3 | 51.1 |
| + SFT + OPD | 31.2 3.5 | 42.4 6.1 | 74.0 19.3 | 62.5 5.0 | 52.5 8.4 |
| Qwen3-VL-4B | 29.0 | 35.2 | 60.7 | 60.9 | 46.5 |
| + SFT | 30.3 | 43.2 | 72.7 | 62.7 | 52.2 |
| + SFT + OPD | 31.0 2.0 | 43.5 8.3 | 72.7 12.0 | 65.7 4.8 | 53.2 6.7 |
| Qwen3-VL-8B | 30.8 | 40.3 | 52.7 | 43.6 | 41.9 |
| + SFT | 31.0 | 43.8 | 72.7 | 58.3 | 51.5 |
| + SFT + OPD | 32.2 1.4 | 44.3 4.0 | 74.0 21.3 | 59.5 15.9 | 52.5 10.6 |
3.4 Ablation Studies
Contribution of Local-Transition Supervision.
For Qwen2.5-VL-7B, external-data-only SFT raises the four-benchmark average from 35.1 to 56.0; adding local interaction supervision reaches 58.4. The controlled incremental gain from L1–L2 data is 2.4 points, considerably smaller than the complete pipeline’s improvement over the untouched backbone.
Contribution of Privileged Process Supervision.
Starting from the same full SFT checkpoint, answer-only GRPO reaches 59.2 overall and OPD reaches 60.1. OPD improves route planning from 28.9 to 34.0 and camera displacement from 33.5 to 36.6 relative to GRPO. This comparison matches data, prompts, rollout groups, rewards, and optimization; it isolates the addition of privileged context together with the process objective.
| Training stage | VSI | MindCube | VSTI | SPBench-MV | Overall | |||
|---|---|---|---|---|---|---|---|---|
| Rel. Dist. | Route | Avg. | Avg. | Displ. | Avg. | Avg. | Avg. | |
| Base | 36.6 | 30.4 | 37.9 | 29.3 | 8.6 | 39.0 | 34.0 | 35.1 |
| External-data SFT | 48.8 | 30.4 | 46.2 | 70.8 | 30.1 | 45.3 | 61.6 | 56.0 |
| Full SFT | 49.9 | 29.4 | 48.2 | 72.2 | 32.6 | 47.5 | 65.5 | 58.4 |
| Full SFT + GRPO | 50.3 | 28.9 | 48.1 | 73.1 | 33.5 | 48.6 | 67.0 | 59.2 |
| Full SFT + OPD | 50.6 14.0 | 34.0 3.6 | 49.0 11.1 | 74.3 45.0 | 36.6 28.0 | 49.5 10.5 | 67.6 33.6 | 60.1 25.0 |
Progressive Transition Curriculum.
The separate curriculum diagnostic compares L1-only, L2-only, and sequential L1→L2 training. The sequential version is the only one to improve all four held-out benchmarks and raises their average from 44.1 to 49.8. This supports complementary local supervision; these subset runs are distinct from the full SFT mixture.
3.5 Process and Temporal Analysis
During training, answer rewards rise while measured process divergence falls. These trends accompany improved scores but do not establish faithful reasoning. Frame shuffling after training hurts the 3B/7B models by 6.6/7.0 points, versus 0.9/0.8 before training, consistent with greater use of order. Increasing route-planning input from 16 to 64 frames yields only a modest OPD gain, 33.4→34.4.
A frozen SFT model scores 32.6 on 839 displacement questions without a trace, 35.4 with four ordered descriptions, 33.8 with those descriptions shuffled, and 35.3 with eight ordered descriptions. The comparison controls the original frames and questions and favors temporal organization over extra segmentation. It supplies descriptions at inference and must not be mistaken for ordinary student evaluation.
3.6 Closed-Loop Interaction
Closed-loop tests compare the base and trained 7B model under matched interaction protocols [14] [18]. On WalkerBench Standard’s 100 tasks, success rises from 7% to 14%; distance and visibility remain at zero. Mean actions over all tasks fall from 17.71 to 6.52, but this alone cannot establish more efficient successful behavior because failed episodes are included.
On eight ESI-Bench categories, each with 30 matched questions, mean success increases from 32.9% to 41.2%. Steps among correct episodes decline from 12.34 to 10.55; that denominator differs from WalkerBench. The appendix’s broader 27-category diagnostic is less favorable: macro success rises from 26.3% to 31.1%, with several regressions and two categories excluded for lack of valid paired runs.
| Model | Angle | Dist. | Height | Nav. | Vis. | Overall | Steps |
|---|---|---|---|---|---|---|---|
| Base | 5.0 | 0.0 | 25.0 | 5.0 | 0.0 | 7.0 | 17.71 |
| Spatial-Interactor | 20.0 | 0.0 | 40.0 | 10.0 | 0.0 | 14.0 | 6.52 |
| Model | Rigid | Stability | View | Occl. | Distance | Agent Obs. | Unobs. | Act. Order | Overall | Steps |
|---|---|---|---|---|---|---|---|---|---|---|
| Base | 33.3 | 26.7 | 56.7 | 23.3 | 50.0 | 20.0 | 26.7 | 26.7 | 32.9 | 12.34 |
| Spatial-Interactor | 40.0 | 36.7 | 63.3 | 33.3 | 56.7 | 33.3 | 33.3 | 33.3 | 41.2 | 10.55 |
![[Uncaptioned image]](https://storage.literature.yihao.one/media/sha256/85/853b590c0ce76ff02977a5bc8c621408699df77e66b2303092bb974eb7af4522.webp)
3.7 Qualitative Analysis
Two selected cases illustrate the intended behavior: recognizing camera rotation through a shared cardboard-box anchor, and distinguishing rightward translation from forward motion using landmark parallax across a longer video. The associated attention maps and explanations illustrate possible evidence use; they do not establish a causal account of the model’s reasoning.
4 Conclusion
The study supports local interaction supervision followed by privileged process training as a way to improve spatial question answering and some interactive tasks. Its strongest controlled evidence is the incremental improvement over matched SFT and GRPO baselines. Closed-loop success remains low or uneven, the broad ESI analysis uses small selected paired subsets, and generated motion descriptions can be imperfect. Fixed temporal segmentation and limited real interaction diversity remain targets for future work.
Appendix A Related Work
A.1 Evaluating Spatial Reasoning in VLMs
Spatial evaluation has expanded from single-view relations to multi-view correspondence, temporal changes, and embodied actions [43] [27] [24] [45] [39] [18] [16] [29]. Together these tests ask whether a model can maintain spatial consistency as observations accumulate, rather than merely recognize a relation in one image.
A.2 Methods for Improving Spatial Reasoning
Prior improvements include spatial instruction datasets, explicit geometry or world-model representations, and intermediate textual descriptions [5] [25] [33] [3] [44] [11] [41] [28] [13] [4] [50] [48] [38] [20] [7]. Spatial-Interactor emphasizes observed action-conditioned transitions as a complementary supervision source.
A.3 On-Policy Distillation
On-policy distillation trains on prefixes sampled by the current student. Related work explores self-distillation and privileged visual or textual context [51] [21] [19] [47] [26]. Here, the additional teacher input is an ordered description of the same video trajectory.
Appendix B Dataset Construction
B.1 Data Sources and Interaction Records
Simulated and Real Sources
Simulation sources are AI2-THOR, ProcTHOR, HSSD, Replica, and SIMS-V. Real camera trajectories come from ScanNet, ScanNet++, MultiScan, ARKitScenes, and RoomTour3D; BridgeData V2 supplies robot end-effector and gripper states [23] [10] [22] [36] [3] [8] [46] [30] [9] [17] [37].
Interaction Records
Each local record connects observations before and after an executed or measured action. Complete sequences remain available for longer tasks. Recorded actions, poses, state changes, or trajectories determine the target before any question is written.
B.2 Ground-Truth Derivation and QA Instantiation
Construction Overview
The retained algorithm converts simulated actions or measured trajectories into canonical targets, rejects unreliable candidates, instantiates deterministic questions, and verifies that the required evidence is visible.
Algorithm 1 Construction of a verifiable interaction QA record.
- Input: a controllable simulated scene or an annotated real trajectory, together with a task-template bank.
- Execute valid simulated actions or normalize real camera and robot trajectories.
- Derive a canonical target directly from the recorded action, state, or geometry.
- Reject failed, weak, discontinuous, ambiguous, or non-unique interactions.
- Render the ordered media, question, options, and answer from using a deterministic template.
- Verify visual answerability, then organize accepted records by source, task, answer, and curriculum level.
Ground-Truth Derivation
Simulator logs provide camera and object changes directly. Real camera poses are normalized to a shared rigid-transform convention; camera centers and relative motion determine signed translation and rotation. Object tracks and robot state changes supply corresponding local targets. Long-horizon endpoint displacement measures horizontal start-to-end separation, while path length sums successive horizontal movements. Endpoint bearing is relative to the initial view when reliable, or the first reliable motion direction for RoomTour3D; it is not final camera yaw.
Equations 3–4
| (3) |
| (4) |
Canonical Records and QA Instantiation
A canonical record stores ordered observations, actions, the derived target, and geometry/quality metadata. Deterministic templates generate action inference, view matching, metric trajectory, temporal ordering, and reverse-path questions. Distractors alter directions or orders. Displayed numeric options may be quantized, while exact measurements remain in metadata.
Equation 5
| (5) |

B.3 Quality Filtering
Filtering removes failed actions, weak or discontinuous motion, ambiguous answers, invalid geometry, and inconsistent action sequences. Cross-view questions require persistent visible anchors. A frozen VLM checks that the final rendered question is answerable from the images and options; it can reject records but cannot change targets. Splits group examples by source scene or video.
B.4 Curriculum Organization and Dataset Statistics
Three-Level Task Taxonomy
L1 includes object-state changes and robot operations, including multi-step operations. L2 covers ego-motion, cross-view localization, parallax, and perspective taking. L3 covers complete camera paths, metrics, turns, revisits, and reversal. Local tasks vary which component of the observed transition the model must infer.



Final Composition and Numeric Targets
LSI-108K contains 107,518 records: 15,109 L1, 69,487 L2, and 22,922 L3. Independently, 55,338 come from simulation, 42,180 from real camera trajectories, and 10,000 from real robot trajectories. These are full-dataset totals, not the smaller training subsets. The distribution plots report geometry-derived targets where the needed metadata exist.
| Curriculum level | Interaction source | ||||
|---|---|---|---|---|---|
| Level | Samples | Fraction | Source | Samples | Fraction |
| L1 | 15,109 | 14.1% | Simulated interactions | 55,338 | 51.5% |
| L2 | 69,487 | 64.6% | Real camera trajectories | 42,180 | 39.2% |
| L3 | 22,922 | 21.3% | Real robot trajectories | 10,000 | 9.3% |
| Total | 107,518 | 100.0% | Total | 107,518 | 100.0% |
Appendix C Method and Training Details
C.1 Offline Privileged Trace Construction
Frozen Visual-Language Annotator and Prompt
The offline annotator sees 32 chronological frames, split into fixed intervals 01–08, 09–16, 17–24, and 25–32. It writes roughly 50 words per interval about visible anchors, changes, and cautious qualitative camera motion. It receives no downstream question, answer, reward target, or task metadata. Failed interval-coverage checks trigger regeneration.

Trace Assembly and Validation
The four descriptions retain their wording and temporal boundaries and are reused for every question about that video. Checks reject hidden-target references and unsupported metric claims. Only the teacher receives the trace, inserted between video and question; geometry-derived answers are used for rewards, never added to the trace.

C.2 On-Policy Same-Prefix Distillation
Student Rollouts
Each student prompt requests a 90–120-word reasoning paragraph followed by a compact answer. A fixed reasoning prefill starts each response, and eight responses are sampled per prompt. The explicit answer boundary identifies which tokens may receive the process loss.

Privileged Teacher and Process Loss
At each update, the teacher is the frozen pre-update policy, not an exponential moving average. It scores the student’s sampled prefixes with the extra trace. Both distributions are renormalized on the teacher’s top 100 non-special tokens at temperature 1.0. Only reasoning positions contribute; the final answer, padding, and special tokens are masked.
Equations 6–7
| (6) | ||||
| (7) |
C.3 Verifiable Reward and Joint Objective
Reward is 90% task correctness and 10% valid reasoning/answer format. Task correctness is exact match, mean field accuracy, or a clipped relative numeric-error reward, depending on the question. Rewards are standardized within each group of eight rollouts for the clipped GRPO objective. A distinct frozen SFT reference supplies the KL penalty with coefficient 0.01.
Equations 8–10
| (8) |
| (9) |
| (10) |
Process-Loss Schedule
The process-loss weight is 0.05 for the first 100 optimization steps, then falls linearly to zero over the next 1,100. Privileged guidance therefore acts mainly during early and middle training.
C.4 Training Configurations
Optimization Hyperparameters
Both stages run one epoch in BF16 with AdamW and a frozen visual encoder. SFT uses a peak learning rate of 10⁻⁵, 3% warmup, cosine decay, and batch 64. GRPO/OPD use 5×10⁻⁷, FSDP full sharding, 16 prompts per batch, eight responses per prompt, sampling temperature 0.8 and top-p 0.9. Inputs contain 32 frames; prompt/response limits are 4,096/384 tokens.
Compute Environment
Training uses one Linux node with eight NVIDIA B200 GPUs. SFT runs in LLaMA-Factory; on-policy training uses EasyR1/verl and vLLM rollouts. The paper reports released entrypoints, reward functions, process masks, configurations, and an environment manifest.
Appendix D Additional Analysis
D.1 Trace-Assisted Discovery Diagnostic
The trace-assisted discovery diagnostic inserts four question-independent segment descriptions directly into the inference prompt. It asks the model to distinguish translation from rotation and endpoint displacement from accumulated path length. This is an auxiliary-context experiment, separate from the student’s normal trace-free inference.

Trace Perturbation Protocol
The order/segmentation diagnostic uses a frozen Qwen2.5-VL-7B SFT model on all 839 camera-displacement questions from 310 videos. No trace, four ordered descriptions, those same descriptions shuffled, and eight ordered descriptions share the same 32 frames, question, deterministic decoding, and official parser. Trace generation cannot see the reference answer.
D.2 Expanded ESI-Bench Breakdown
The expanded ESI-Bench comparison uses 30 matched questions per category selected from completed paired runs, across 27 categories. Macro success improves from 26.3% to 31.1%, but connectivity, inclined-plane tasks, long-term navigation, and traversable passage regress. Category ambiguity and merged observation remain at zero. Deformable and liquid-volume categories are excluded because valid paired runs were unavailable; these selected subsets are a diagnostic, not a complete benchmark evaluation.
| Category | Base | Ours | Category | Base | Ours | Category | Base | Ours |
|---|---|---|---|---|---|---|---|---|
| Action Order Inference | 26.7 | 33.3 | Inclined Plane | 16.7 | 6.7 | Rigid Containment | 33.3 | 40.0 |
| Agent Observation | 20.0 | 33.3 | Linear Alignment | 50.0 | 60.0 | Spatial Distance | 50.0 | 56.7 |
| Category Ambiguity | 0.0 | 0.0 | Long-Term Navigation | 20.0 | 6.7 | Spatial Relations | 50.0 | 60.0 |
| Connectivity | 53.3 | 40.0 | Material Transparency | 43.3 | 46.7 | Spatial Segmentation | 3.3 | 13.3 |
| Correspondence | 20.0 | 26.7 | Merged Observation | 0.0 | 0.0 | Stacking & Stability | 26.7 | 36.7 |
| Counting w/ Occlusion | 0.0 | 6.7 | Partial Occlusion | 23.3 | 33.3 | Structural Enclosure | 3.3 | 3.3 |
| Dimensional Size | 0.0 | 20.0 | Physical Contact | 56.7 | 66.7 | Traversable Passage | 40.0 | 20.0 |
| Geometric Configuration | 6.7 | 26.7 | Reflection Authoring | 50.0 | 63.3 | Unobserved Change | 26.7 | 33.3 |
| Illumination Variability | 0.0 | 3.3 | Regional Boundary | 33.3 | 40.0 | View Hallucination | 56.7 | 63.3 |
| Macro Average | 26.3 | 31.1 | ||||||
Appendix E Complete Visual Task Gallery
All 35 task examples are retained with their original observations, questions, options, and ground truths. E01–E10 illustrate object and robot transitions; E11–E27 cover camera motion and viewpoint reasoning; E28 returns to a multi-step robot operation; E29–E35 cover complete paths, metrics, turns, and revisits. The gallery illustrates task construction, not model predictions.


































