Proximity Weighted Losses for Physical Interaction Modeling

31 concepts19 papers

When deep learning models generate or track physical interactions—whether it is two people swing dancing, a hand slicing a lemon, or furniture arranged in a living room—standard spatial loss functions frequently fail.

The root of this failure lies in the reliance on absolute coordinate systems. If a generative model predicts a hand's position with a minor two-centimeter error, a standard Mean Squared Error (MSE) loss penalizes that error identically regardless of context. If the hand is moving through mid-air, a two-centimeter shift is visually imperceptible. However, if the hand is supposed to be shaking another person's hand or gripping a cup, that same two-centimeter error results in a glaring artifact: a "near-miss" floating contact or a physically impossible interpenetration.

To solve this, a prominent research direction focuses on proximity-weighted and relative spatial losses. By shifting the optimization objective from absolute global coordinates to the relative distances between interacting entities, these methods ensure that models prioritize precision exactly where physical contact occurs.


The Illusion of Absolute Accuracy

Standard training objectives evaluate joints, vertices, or objects in isolation. As highlighted by Liang Xu et al. in their work on ReGenNet, minimizing absolute global coordinate errors does not guarantee that the synthesized movements will remain synchronized or physically grounded to an interaction partner.

When humans evaluate an interaction, our perception of realism is highly sensitive to relative spatial relationships rather than global positioning. As Xiyan Xu et al. argue in MoReact, prioritizing the relative distance between interacting joints over their absolute coordinates is the single most effective factor in preventing visual artifacts like missed contacts or body penetrations.


The Mechanics of Proximity Weighting

The core mechanism of a proximity-weighted loss function is an adaptive scaling factor, often governed by exponential decay. The closer two interacting joints or surfaces get to one another, the higher the penalty becomes for any spatial deviation between them.

$$\text{Loss} = \sum_{i,j} w(d_{ij}) \cdot \mathcal{L}_{\text{spatial}}(x_i, x_j)$$

Where the weight $w(d_{ij})$ increases sharply as the distance $d_{ij}$ between entity $i$ and entity $j$ approaches zero.

Weight (Penalty)
  ^
  |  *
  |   *
  |    *
  |     *
  |       *
  |          *
  +------------------------> Distance between interacting entities
  0 (Contact Point)

This dynamic scaling forces the network to focus its capacity on high-stakes contact zones. For example, in REMOS, A. Ghosh et al. employ this exponential decay strategy to synthesize reactive two-person motions, ensuring that synchronization is at its sharpest during high-contact moments like martial arts or dancing. Similarly, Qingxuan Wu et al. demonstrate in Text2Interact that weighting joint pairs more heavily when they are physically close allows generative diffusion models to accurately capture the defining features of intimate human-to-human interactions like handshakes or embraces.

This principle is equally powerful for human-object interactions. In Person in Place, ChangHee Yang et al. incorporate a distance-based penalty into a diffusion loss function that increases the penalty for joints that wander away from their target interaction sites. This constraint acts as a spatial magnet, pulling generated human skeletons toward objects to guarantee natural, physically plausible contact.


Supervising the Interaction Zone in Vision Tasks

Beyond generating 3D poses, proximity-aware supervision has transformed computer vision tasks like video object segmentation and manipulation video generation. In these domains, the loss function is directed to supervise the "interaction zone"—the precise boundary pixels where a manipulator touches an object.

In HOIST-Former, Supreeth Narasimhaswamy et al. demonstrate that explicitly supervising pixels at these contact points trains the model to cleanly delineate the boundaries between a hand and a held item, even under severe occlusion.

Proximity weighting also acts as a powerful noise filter. In ActionVOS, Liangyang Ouyang et al. introduce a weighted focal loss that assigns higher importance to pixels where an object is simultaneously referenced in text and in direct physical contact with a hand. This prevents the model from being distracted by background clutter that happens to be near the object. When generating video sequences of complex manipulations, Tianxiao Chen et al. show in MIMIC that prioritizing these interaction zones eliminates common generative artifacts like ghosting or blurring around the hands, preserving sharp contact dynamics over time.


Enforcing Relative Constraints in Structural Assemblies

The utility of relative position losses extends to macro-level spatial arrangements and structural assemblies, where physical overlap is the primary failure mode.

  • Furniture Arrangement: In Graph Neural Network for Generative Furniture Arrangement, Miao Yang et al. utilize relative position loss functions to penalize deviations in inter-object distances compared to ground-truth configurations, significantly reducing physical overlaps and respecting real-world functional proximity.
  • Orthodontics: A parallel approach is demonstrated by Changsong Lei et al. in Automatic Tooth Arrangement..., where a relative position loss prevents teeth from intersecting or overlapping during automated orthodontic alignment.
  • Kinematic Chains and Multi-Object Interactions: For full-body movements, Ling-An Zeng et al. introduce auxiliary losses in ChainHOI that penalize the distance between human joints and object surfaces to eliminate floating artifacts. When scaling up to multiple objects, Xintao Lv et al. show in HIMO that explicitly modeling relative object-to-object distances maintains physical plausibility across complex multi-contact timelines.

Furthermore, relative constraints allow models to preserve the semantic meaning of interactions across different physical scales. In Semantics2Hands, Zijie Ye et al. leverage a proximity-weighted reconstruction strategy to ensure that fine-grained gestures (such as fingers touching to form a pinch) are preserved when retargeting hand motions between avatars with vastly different hand proportions. Similarly, Lan Xu et al. utilize specialized interaction losses in HandDiffuse to penalize joint penetration and enforce plausible hand shapes during complex two-hand interactions.


Rethinking Evaluation via Relative Metrics

If absolute coordinate losses fail to train effective interaction models, standard absolute metrics fail to evaluate them accurately. Traditional pose estimation metrics align and evaluate objects or hands independently in a global coordinate space, completely missing whether the interaction itself is physically sound.

To address this evaluation blind spot, Gyeongsik Moon et al. introduce a relative vertex error metric in A Dataset of Relighted 3D Interacting Hands. Instead of measuring how far a predicted hand is from its global ground truth, this metric measures the accuracy of the distance between the two interacting hands. By evaluating spatial relationships directly, relative metrics successfully capture the subtle nuances of physical contact and occlusion that traditional independent metrics overlook, providing a benchmark that aligns with physical reality.

Key Foundational Work

Graph Neural Network for Generative Furniture Arrangement (2021) — contributed 2 concept elements to this direction.
PLACE: Proximity Learning of Articulation and Contact in 3D Environments (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_proximity_weighted_interaction_loss_f647d3f4,
  title = {proximity-weighted interaction loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proximity-weighted-interaction-loss/rdr_f647d3f4f42448bd9d45ac6adb947dbd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_proximity_weighted_interaction_loss_f647d3f4_v1,
  title = {proximity-weighted interaction loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proximity-weighted-interaction-loss/rdr_f647d3f4f42448bd9d45ac6adb947dbd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake