WAMotor inference acceleration for Faster-WAM
The WAMotor authors published a speculative inference implementation for Faster-WAM. It reuses denoising velocities without training a separate draft model and was evaluated on LIBERO, RoboTwin 2.0 and LIBERO-Plus.
Previously reported
The study uses this model’s standard flow-matching inference as its baseline.
This announcement
The authors used an RTX 5090 and measured inference latency per action chunk and mean task success. Each row compares standard inference with WAMotor. LIBERO (40 tasks × 50 trials) — 275.85 → 90.10 ms (3.06×), 98.60% → 99.20% RoboTwin 2.0 (50 tasks × 100 trials in each of clean and randomized settings) — 282.09 → 98.70 ms (2.86×), 92.45% → 92.05% LIBERO-Plus (10,030 instances × 3 trials across seven types of shift) — 282.62 → 90.33 ms (3.13×), 74.52% → 75.05%
Faster-WAM caches future context in one pass and uses sparse video–action interactions during speculative denoising. The WAMotor implementation defaults to 10 denoising steps, BF16 and Shift-5. Speculation starts after one Euler update, uses a verification threshold of 0.01 calibrated to the median step size and caps the window at the denoising step count. Enable it with --speculative. The LIBERO-Plus launcher defaults to a 15% task sample, seed 42 and one trial per instance. The numbers above come from the full study protocol. Upstream lists step 021700 for LIBERO and LIBERO-Plus, and step 029355 for RoboTwin 2.0. The WAMotor paper gives no exact checkpoint hashes. The repository provides a 2-camera LIBERO task config (also used for LIBERO-Plus) and a 3-camera RoboTwin task config. Their mode is one_pass_future_cache.
