⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

KAM-WM: Kinematic Affordance Maps from Latent World Models for Robot Manipulation

arXiv 2026 40.1 method

TLDR

KAM-WM extracts kinematic affordance maps from a frozen latent video world model to guide robot manipulation policies, achieving high success on LIBERO and RoboTwin2.0.

Reasoning

The paper presents a novel method that leverages a frozen video world model to provide directional interaction cues without rollout or fine-tuning, demonstrating strong empirical results. However, the approach is evaluated only on two specific benchmarks and relies on a pre-trained backbone, which may limit generalizability and adaptability to new tasks.

Read-first score

Read-first score 40.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 36.

Recency 6%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%
70

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,result

Topical relevance 29%
51.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Citation impact 18%
0

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%
0

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 284.

Keyword Scores

world model
9
video world model
9
generative world model
6
world dynamics prediction
5
interactive world model
3
world simulator
2
model-based reinforcement learning world model
2

Deep Analysis

Innovations

  • Extraction of kinematic affordance maps (directional interaction cues) from a frozen latent video world model without rollout or fine-tuning
  • Interpretation of single-step latent velocity from a Flow Matching image-to-video backbone as a Kinematic Affordance Map (KAM)
  • Compression of KAM via a lightweight Perceiver to condition a diffusion policy alongside RGB and proprioception
  • Demonstration that directional information beyond spatial localization improves few-shot manipulation performance

Methodology

KAM-WM queries a frozen Flow Matching image-to-video model once, interprets its latent velocity as a Kinematic Affordance Map encoding task-conditioned interaction regions and coarse motion structure. A Perceiver compresses this map into tokens that condition a diffusion policy together with RGB observations and proprioception. The framework is evaluated on LIBERO and RoboTwin2.0 benchmarks.

Key Results

KAM-WM achieves 90.6% average success on LIBERO, 65.7% on RoboTwin2.0 Easy, and 22.4% on RoboTwin2.0 Hard. Controlled comparisons against a zero-order mask prior show that directional information contributes to performance gains beyond spatial localization.

Limitations

  • Provides only coarse directional cues and coarse motion structure, not fine-grained motion
  • Low success rate on RoboTwin2.0 Hard setting (22.4%)
  • Evaluated only on LIBERO and RoboTwin2.0; generalization to other tasks and domains unclear

Resources

Paper

Tags