acceptodds
Under review as a conference paper at ICLR 2027

SLoHa: A 4D Human Locomotion Dataset of Scene-constrained Large-object Handling via a Democratic Data Acquisition Solution

Abstract

Carrying large objects through cluttered environments is an important capability for robots in daily assistance, logistics, and workplaces, requiring coordinated whole-body motion, object control, and obstacle avoidance. Yet learning this capability remains underexplored, partly due to the scarcity of suitable training data. Human demonstrations provide a valuable reference for robots to learn such coordinated behaviors. However, existing datasets primarily capture small-object interactions in unconstrained environments and often rely on expensive motion-capture systems that are difficult to scale, limiting the availability of scene-aware large-object demonstrations and hindering data-driven learning of this capability. To bridge this gap, we present SLoHa, a dataset for scene-constrained large-object handling, together with a democratic data acquisition solution that makes such demonstrations more accessible to collect. Using four exocentric cameras, modern reconstruction priors, and sparse human-assisted alignment, our pipeline recovers human motion, object geometry and 6D pose trajectories, and scene constraints in a shared coordinate system without specialized MoCap suits or object markers. SLoHa contains 21.4 hours of demonstrations involving 150 large objects and 435 obstacle layouts. We present a scene-conditioned large object handling benchmark. Experiments show that our dataset improves constraint-compliant motion generation, validating its effectiveness for studying scene-aware interaction and its potential use in embodied applications.

Primary area:
datasets and benchmarks
Keywords:
4D Human Locomotion Dataset, Large-object Handling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zQniOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SLoHa captures 21.4 hours of large-object carrying using four cameras, covering 152 objects and 435 modular obstacle layouts in two rooms. It evaluates scene-conditioned motion generation and humanoid motion tracking.

Critical strengths

The dataset captures a useful interaction regime. Large-object carrying couples foot placement, body posture, object orientation, and environmental clearance. SLoHa records these adjustments around physical obstacles, including narrow passages and overhead barriers (Section 3). This coverage gives scene-conditioned models information that isolated human–object interaction data cannot provide.

The matched generation experiment demonstrates useful scene information. The authors hold CHOIS, training data, and task inputs fixed while adding scene occupancy (Section 4.1). The fraction of penetrating object vertices falls from 27% to 18%, and human vertices from 16% to 13% (Table 3). This isolates a measurable benefit of the reconstructed scenes.

Critical weaknesses

Geometric label accuracy lacks quantitative validation. Appendix A.4 provides ten paired MoCap sequences but relies on visual comparisons; smoothness and contact statistics do not measure body or object pose error. This matters because the task depends on precise relative human–object–obstacle geometry.

The scene-diversity claim conflates layouts with environments. At least some visual difference must be shown on top of the box collison. Table 1 counts 435 modular box-obstacle layouts as "scenes," although collection used only two rooms (Section 3.1; Appendix A.1). Five layout types account for the data, and 73% are straight or corner passages (Figure 4). This measures local constraint variation, not breadth of physical environments, and makes comparisons potentially misleading. The sequence-level 90/10 split does not test held-out layouts or rooms; Section 4.3 supplies only qualitative obstacle-replacement examples. IF author can justify that, i am happy to discuss possible improvement of score.

Questions for authors

Can the authors report held-out-layout or held-out-room results to support generalization beyond the 435 box arrangements?

What are the body-joint and object-pose errors before and after alignment?

1 comment

Thank you for the clarifications and the additional reconstruction results. I appreciate that participants navigated real obstacles and adjusted both their body motion and the object pose. My main concern is what these demonstrations add beyond existing human–object motion data combined with scene geometry and collision constraints. The scene-conditioning ablation shows that geometry helps, but it does not isolate the benefit of collecting demonstrations around physical obstacles. The HUMOTO comparison is a useful first step, but it is qualitative and uses training sets of different sizes. The benchmark also mainly tests box-lined corridors. These layouts capture the challenge of carrying large objects through tight spaces, but the results do not show whether the model handles irregular obstacles in more cluttered rooms. Unless more concrete proof is shown, I am keeping my original score.

Reviewer_DftoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SLoHa, a 4D human-scene-object dataset for large-object carrying under scene constraints. It uses four GoPro cameras and VGGT, ReconViaGen, SAM2, FoundationPose, and SMPL-X to reconstruct human motion, object geometry and 6D pose, and box layouts. The dataset has 21.4 hours, 152 objects, and 435 layouts. The paper validates the data with a CHOIS scene-conditioning benchmark, unseen-object tests, and G1 humanoid retargeting.

Critical strengths
  1. The dataset fills a real gap: prior work either has large objects without scene constraints, or scene constraints without large objects.
  2. The capture pipeline is cheap and needs no MoCap suits or object markers: four GoPros cost about $1500, and no MoCap suits, object markers, or manual CAD are needed.
  3. The scene-conditioning ablation is clean: the same CHOIS backbone and data split, with only the scene input changed; Table 3 shows lower penetration.
Critical weaknesses
  1. The central design choice lacks a direct empirical comparison with post-hoc scene synthesis. The paper physically builds box scenes and then reconstructs them, and gives a conceptual reason (p.2, p.14). MIME, SUMMON, and Pose2Room generate or infer scene arrangements compatible with human motion; the paper does not discuss this line. However, these methods do not themselves generate the scene-induced carrying motions targeted by SLoHa, so a comparison would need to account for this difference. For such an alternative, deriving test-time scene conditions from the full target motion could inflate the scene-conditioning gain.
  2. The human cost is not fully reported. The paper explicitly uses human-in-the-loop alignment. About 30% of sequences need pose-alignment annotation at about 30 seconds per corrected sequence (p.2; Sec. 3.3; Appendix A.1, A.2; Sec. 3.4). But for labor cost, it reports the annotation rate and time and the camera cost, not total person-hours. It gives no annotator count or agreement. The person-hours for mask verification, visual inspection, calibration, and capture are also not reported.
  3. "Large object" is not quantitatively characterized, and the contribution of object size is not isolated. The paper gives examples and a two-handed carrying criterion, but no size or weight threshold and no per-object dimensions. Table 1 lists total object counts without a common large-object criterion across datasets. The paper gives a mechanism argument (swept volume), but Table 5 (Appendix A.5) compares SLoHa with OMOMO, and the two datasets differ in data scale, object count, and capture setup. This comparison shows a dataset-level gain but does not isolate the effect of object size.
  4. Reconstruction accuracy is not quantitatively validated. The only accuracy check is a qualitative Nokov MoCap overlay on 10 sequences (Fig. 9, Appendix A.4), and the paper itself calls it qualitative. The paper reports proxies such as foot sliding, jerk, and penetration, but no MPJPE, object pose error, or box-geometry error. It also has no camera-based baseline or component ablation.
  5. The appendix reports real-robot experimental results, while the limitations section describes closed-loop deployment on physical humanoid robots as a direction for future exploration. The relationship between these results and the stated limitation is unclear. I also suggest moving the real-robot results to the main text.
Questions for authors
  1. Can you provide direct empirical evidence for physically capturing scene-constrained motions rather than pairing existing motions with synthesized layouts? If a motion-to-scene baseline is used, how would you prevent its test-time scene conditions from revealing the full target motion?
  2. What is the total number of person-hours required for the full data acquisition pipeline?
  3. Beyond the two-handed carrying criterion, were any quantitative size or weight thresholds used, and what are the size and weight ranges of the included objects?
  4. Can you report quantitative human and object reconstruction errors against Nokov, and box-geometry errors against the measured layouts?
Reviewer_ZqyhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SLoHa, a 21.4-hour markerless 4D human-object interaction dataset of humans carrying 152 large objects through scene-constrained corridors. The data is reconstructed using multiple GoPro cameras and pretrained priors for scene, human, and object tracking. Experiments demonstrate downstream utility across scene-conditioned locomotion synthesis, motion forecasting, and humanoid skill learning.

Critical strengths

The critical strength of the paper is its problem formation and dataset scale. The paper identifies a good and practical gap in 4D HOI research by coupling large object transportation with 3D environmental constraints. It is also scalable and low cost, showing some reasonable quality compared to other datasets.

Critical weaknesses

I think the major question is that the author didn't fully isolate dataset value from conditioning architecture. The author only compare a model trained on SLoHa with and without scene conditioning. This should only proves that feeding obstacle geometry into a diffusion model reduces obstacle penetration, rather than proving the value of SLoHa's captured motions. I would like to see a comparison on an existing unconstrained dataset augmented with synthetic bounding-box obstacles to prove that captured avoidance adapatations outperforms the synthetic ones.

Questions for authors
  1. Narrow straight pathways account for 56% of the dataset. How does this distribution imbalance affect motion generation performance across specific obstacle types, and can the authors report a per-layout breakdown of the penetration metrics in Table 3?
  2. The hand motion seems not being considered in the paper and most of the showed cases require two hands or arms to hold an object. Can author show cases where the carrying pose are not two hand "hugging-like" poses?
  3. Please provide the simulation setup in more detail, including object mass, gravity, and friction coefficients.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.