AnyStep-WAM Cuts Denoising Steps by Up to 85% in World Action Models Without Losing Success Rate
Rui Wang · hf · 2026-09-30
Researchers introduce AnyStep World Action Model (AnyStep-WAM), a general framework for tunable-budget prediction and scene-dependent compute allocation in world-action models (WAMs).
Motivation: Manipulation tasks contain action chunks with varying sensitivity to generation errors, yet WAMs typically use a fixed number of denoising steps for all actions.
Method:
- Budget-aligned teacher-trajectory distillation trains interval-conditioned flow maps with frozen-teacher transitions and shared low-rank adapters, supporting one-step to multi-step action generation
- A lightweight risk-benefit scheduler predicts teacher-curvature-based difficulty and student-teacher fidelity from a single one-step preview, selecting the smallest budget that meets risk-adaptive fidelity requirements
Results (on Motus, FastWAM, and LingBotVA using RoboTwin 2.0):
- Average denoising steps reduced by 60.2%, 49.8%, and 85.28% with baseline task success rates maintained
- Under a one-step budget, AnyStep training boosts success rates by 7.07%, 12.08%, and 8.94% respectively
- Validated on six real-world manipulation tasks
More from Embodied
- Simify: training-free real-to-sim framework solves robot spatial reasoning in seconds — Ed__Johns · 2026-09-30
- ZeroBot trains robot manipulation from scratch in 119 seconds with 87% success using generative real2sim — Ed__Johns · 2026-09-30
- LadderMan, zero-shot sim-to-real humanoid ladder climbing, wins CoRL 2026 Spotlight — yuewang314 · 2026-09-30
- Reddit users warm to Goose, a non-humanoid home robot that fetches and grabs — fhpapa · 2026-09-30
- Amazon's StructRL uses verifiable intermediate rewards to improve long-horizon robot manipulation — amazon · 2026-09-30
- Memory price surge threatens on-device AI hardware: BOM tops $300, device may cost $1,500 — stephen280up · 2026-09-30