RegisterWAM adds learnable registers to world action models, hitting 81.6% on LIBERO-Plus and 93% on RoboTwin 2.0
AutoLab-SJTU · hf · 2026-10-15
AutoLab (SJTU), CAS, and Humanoid Robot (Shanghai) study how World Action Models organize action-relevant information, identifying persistent high-norm spatial outliers whose action relevance survives normalization and K/V projection.
- Proposes RegisterWAM: learnable non-spatial registers added to the world model, used directly for action conditioning.
- Registers alone don't fully eliminate outliers; pairing with observation flow-matching (FM) loss supervision suppresses them and concentrates high-norm responses in registers.
- Observation FM supervision improves action learning even without registers, motivating a noised-observation FM objective for scene recovery from corrupted inputs.
- Results: 81.64% on LIBERO-Plus, 93.0% average on RoboTwin 2.0, and better real-world manipulation — without extra embodied pretraining data or foundation models.
More from Embodied
- MimicX refines humanoid tracking with policy-in-the-loop feedback: 25.7% error cut — Shuaijun Liu · 2026-10-15
- Gracia's new world model targets Real2Sim2Real and Physical AI — jonstephens85 · 2026-10-15
- Niantic Spatial built VPS on Pokémon Go imagery; a dev already prototyped it solo — jonstephens85 · 2026-10-15
- RoboParty launches RP1, a fully open-source full-stack humanoid robotics platform — GuanyaShi · 2026-10-15
- Survey: Towards Agentic Robots organizes the field into 17 categories across four domains — Chonghan Qin · 2026-10-15
- HumanoidToolBench benchmarks humanoid tool use from selection to mobile execution on real Unitree G1 — snupilab · 2026-10-15