Ego-Vision World Model for Humanoid Contact Planning

arXiv 25.10 2025 56.3 method, application

TLDR

A framework combining a learned world model with MPC for humanoid contact planning, trained on offline data and deployed on a real robot.

Reasoning

The paper presents a novel integration of a world model with sampling-based MPC for contact-rich humanoid tasks, demonstrating real-world deployment and improved sample efficiency. However, the scope is limited to contact planning and the abstract does not detail comparisons to other world model approaches or ablation studies.

Read-first score

Read-first score 56.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

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

Reproducibility 25%

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

Topical relevance 42%

45.7

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 344.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Combining a learned world model with sampling-based Model Predictive Control (MPC) for humanoid contact planning
Training on a demonstration-free offline dataset to predict future outcomes in a compressed latent space
Using a learned surrogate value function to address sparse contact rewards and sensor noise for dense, robust planning
A single scalable model supporting multiple contact-aware tasks with improved sample efficiency and multi-task capability over on-policy RL

Methodology

The framework integrates a learned world model, trained on an offline dataset without demonstrations, to predict future states in a compressed latent space. A sampling-based Model Predictive Control (MPC) planner uses a learned surrogate value function to handle sparse contact rewards and sensor noise, enabling dense and robust planning. The system is deployed on a physical humanoid using proprioception and ego-centric depth images.

Key Results

The approach achieves improved sample efficiency and multi-task capability compared to on-policy reinforcement learning, and is successfully deployed on a physical humanoid for real-time contact planning tasks including wall support after perturbation, blocking incoming objects, and traversing height-limited arches.

Limitations

Reliance on the quality and coverage of the offline dataset, which may not capture all possible contact scenarios
Potential prediction errors in the learned world model that could affect planning accuracy
Evaluation limited to three specific contact-aware tasks, leaving generalization to broader environments unverified

Resources

Paper Project