Research map / Video / Video generation

LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE

NeurIPS 2026 · Video › Video generation · approach: World-model video generation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LongScape, a hybrid diffusion-autoregressive embodied world model that partitions videos into action-guided variable-length chunks and routes each rollout step to one of four specialized CogVideoX-initialized DiT experts via a context-aware router.

Problem

Long-horizon video generation for embodied manipulation suffers from temporal inconsistency and drift in diffusion models and reduced visual detail in autoregressive models; fixed-length chunks in hybrid models split or merge semantic actions.

Approach

Partition video into variable-length chunks (8 to 32 frames) using gripper state changes and end-effector motion magnitude, train a DiT expert per chunk type initialized from CogVideoX, and use a cross-attention router over text and the current chunk to pick the expert for each autoregressive step.

Key results

• On LIBERO, LongScape reaches PSNR 19.977, LPIPS 0.1231, SSIM 0.7883, FVD 153.72, versus CogVideoX FVD 184.69 and NOVA FVD 346.91.
• On AGIBOT-World, LongScape reaches PSNR 16.493 and FVD 256.16, versus CogVideoX FVD 267.39 and NOVA FVD 273.60.
• The dynamic router achieves 91.4% accuracy (11,606 of 12,700) on a four-class expert selection test set on LIBERO.

Surprising findings

• Fixed 8-frame chunk experts perform worse than longer fixed chunks (32-frame), which the authors attribute to short chunks failing to capture complete actions.

Limitations

• Chunking relies on heuristic thresholds rather than learned partitioning.
• Only single-view inputs; no multi-view or multi-minute generation evaluated.
• Improvements over baselines are modest on AGIBOT-World (e.g., LPIPS 0.3613 vs CogVideoX 0.4028 but PSNR gap small); evaluation uses only four metrics without human or task-success evaluation.

Experimental setup

Datasets: LIBERO AGIBOT-World

Benchmarks: LIBERO AGIBOT-World

Metrics: PSNR LPIPS SSIM FVD

Baselines: CogVideoX NOVA Genie

Builds on: CogVideoX NOVA Mixture-of-Experts Diffusion Transformer

Closest papers by meaning

• RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.89)
• EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation (NeurIPS 2025 · similarity 0.88)
• SAMPO: Scale-wise Autoregression with Motion Prompt for Generative World Models (NeurIPS 2025 · similarity 0.88)
• Trimming the Long-Tail of Visual World Modeling Evaluation (NeurIPS 2026 · similarity 0.88)
• VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · similarity 0.88)
• Empowering World Models with Reflection for Embodied Video Prediction (ICML 2025 · similarity 0.88)