WAVE-0 adds a 257M world model to VLA, lifting real-world success from 38.3% to 73.3%
Haomin Wang · hf · 2026-10-14
WAVE-0 (SJTU, Baidu et al.) bridges VLA policies and world action models:
- Approach: a 257M lightweight world model on a pretrained VLA translates latent action plans into predicted future representations that condition the action expert, unifying latent planning, future prediction and action generation
- Training: trained on ground-truth robot actions with a counterfactual ranking objective — no pixel reconstruction or large-scale video pretraining
- Results: 91.47% on RoboTwin 2.0, 98.7% on LIBERO, and real-world success for π0.5 improved from 38.3% to 73.3%
More from Embodied
- NTU team has Claude write robot control code, hitting 70.24% success in sim — aigclink · 2026-10-14
- Rabbit R1 called 'ahead of its time' as modders plan to hack the once-mocked AI device — SimonBalmain · 2026-10-14
- RSI-Nav: recursive self-improvement for navigation agents hits 68.6% SR on R2R-CE — OpenGVLab · 2026-10-14
- Cepheid decouples scale from structure for cross-family sparse-to-dense motion completion — NYCU · 2026-10-14
- Visor: open-source Android app replaces the Meta AI app on Ray-Ban Meta glasses with your own model — jedisct1 · 2026-10-14
- 304 robotics papers hit arXiv in a single day as the field explodes — DJiafei · 2026-10-14