Research map / Vision-language and multimodal / Robot action and navigation
Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model
John Won, Kyungmin Lee, Huiwon Jang, Dongyoung Kim, Jinwoo Shin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dual-stream MMDiT with decoupled noise (DUST) instead of GR00T-N1.5 | GR00T-N1.5 · RoboCasa (100 demos/task) | average success rate (%): 41.7 -> 50.1 | Improved |
| Use DUST instead of FLARE world modeling | GR00T-N1.5 · GR-1 (1,000 demos/task) | average success rate (%): 36.3 -> 42.0 | Improved |
| Use DUST on real Franka robot tasks | GR00T-N1.5 · Real-world Franka Research 3 | average success rate (%): 46.5 -> 59.9 | Improved |
| Add decoupled (independent per-modality) noise to MMDiT (ablation) | DUST · RoboCasa (100 demos/task) | average success rate (%): 38.2 -> 50.1 | Improved |
| Use dual-stream MMDiT instead of single-stream DiT (ablation) | DUST · RoboCasa (100 demos/task) | average success rate (%): 42.5 -> 50.1 | Improved |
| Pretrain on action-free BridgeV2 video before finetuning | DUST · RoboCasa (100 demos/task) | average success rate (%): 50.1 -> 58.5 | Improved |
| Joint-train with GR-1 and EgoDex human data mixture | DUST · RoboCasa (300 demos/task) | average success rate (%): 58.5 -> 64.4 | Improved |
| Joint-train with GR-1 and EgoDex data mixture | GR00T-N1.5 · RoboCasa (300 demos/task) | average success rate (%): 45.0 -> 46.9 | Improved |
| Increase vision denoising steps from 4 to 64 (asynchronous sampling) | DUST · RoboCasa (1,000 demos/task) | average success rate (%): 66.3 -> 69.7 | Improved |
| Increase vision denoising steps from 4 to 32 | DUST · GR-1 (1,000 demos/task) | average success rate (%): 42.0 -> 47.1 | Improved |
| Increase vision denoising steps from 32 to 64 | DUST · GR-1 (1,000 demos/task) | average success rate (%): 47.1 -> 45.0 | Hurt |
| Set world-modeling loss weight to 0.2 instead of 1.0 (ablation) | DUST · RoboCasa (100 demos/task) | average success rate (%): 50.1 -> 34.3 | Hurt |
Closest papers by meaning
- Unified Diffusion VLA: Vision-Language-Action Model via Joint Discrete Diffusion Diffusion Process (ICLR 2026 · similarity 0.91)
- Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 2025 · similarity 0.89)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · similarity 0.88)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
