RealtimeWAM: One-Step Asynchronous World Action Model Delivers 25x Speedup with <1% Accuracy Loss
NanyangTechnologicalUniversity · hf · 2026-10-06
Researchers from NTU present RealtimeWAM, an efficient World Action Model variant tackling two inference bottlenecks in MoT-based WAMs.
- One-step action generation: Teacher-Anchored Consistency Distillation (TACD) addresses a local-global error gap—low local consistency error alone doesn't guarantee accurate final actions—by adding explicit supervision from the frozen teacher's multi-step rollout endpoint.
- Asynchronous experts: Cross-Expert Wavefront Pipelining (CEWP) overlaps the video and action experts via block-wise KV cache sharing, synchronizing only before the action attention consumes the cache.
Across LIBERO, LIBERO-Plus and RoboTwin benchmarks and model variants, RealtimeWAM maintains near-lossless performance (<1% drop) with large end-to-end speedups (e.g., 25x in simulation on H100). Code and checkpoints are open-sourced.
More from Embodied
- The robotics data business's biggest bottleneck is operational, not technical — paigeinsf · 2026-10-06
- GPT-2-powered robot Samson explores, gets a little confused but has fun — MikePFrank · 2026-10-06
- VC: almost none of real robot deployment data flows back into training — carrycooldude · 2026-10-06
- ACG-Bench Probes Dual-Arm VLA Generalization; AE-VLA Lifts Success from 3% to 21.5% — Zaibin Zhang · 2026-10-06
- ReSteer open-sourced: fixing VLA policies that ignore mid-execution instruction switches — siddkaramcheti · 2026-10-06
- DIY robotic backboard uses computer vision to make every beginner shot go in — TinfoilTricorn · 2026-10-06