Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report
TLDR
World models are a powerful paradigm in AI and robotics, enabling agents to reason about the future by predicting visual observations or compact latent states.
Reasoning
Fallback reasoning generated from available title and abstract metadata: World models are a powerful paradigm in AI and robotics, enabling agents to reason about the future by predicting visual observations or compact latent states. The 1X World Model Challenge introduces an open-source benchmark of real-world humanoid interaction,...
Read-first score
Read-first score 42.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals.
Field roles
Rank sensitivity
Stability: volatile; rank range: 320.
Deep Analysis
Innovations
- Adapting video generation foundation model Wan-2.2 TI2V-5B to video-state-conditioned future frame prediction
- Conditioning video generation on robot states using AdaLN-Zero
- Post-training the model using LoRA
- Training a Spatio-Temporal Transformer from scratch for the compression track
Methodology
The challenge has two tracks: sampling (forecasting future image frames) and compression (predicting future discrete latent codes). For sampling, the authors adapt the video generation foundation model Wan-2.2 TI2V-5B, conditioning on robot states via AdaLN-Zero and post-training with LoRA. For compression, they train a Spatio-Temporal Transformer model from scratch.
Key Results
The model achieves 23.0 dB PSNR in the sampling task and a Top-500 CE of 6.6386 in the compression task, securing 1st place in both challenges.