SmolWAM: A Latent World-Action Model for Robotic Manipulatio
Overview
Source roundup from published reports. Claims below are attributed to their publishers, not independently verified. arXiv Robotics — research abstracts: SmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices. This paper introduces SmolWAM, a latent world-action model designed for robotic manipulation on edge devices. It is derived from the 2B-parameter Cosmos Policy by retaining its pretrained trunk and rebuilding the inferen…
Generated from attributed reports · Updated 35 minutes ago
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · inference speed: 3 other · Basis not reported
“y 21 percentage points in task success rate while achieving approximately three times faster inference.”
Exact source · revision 1Source owner not reported
Reported quantity · GPU-hours: 3.6 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · LIBERO episodes: 400 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · inference time: 57 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · FLOPs per 16-action chunk: 2.3 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · GPU memory: 2.4 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · speed improvement: 11 other · Basis not reported
“runk and rebuilding its inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in one rectified-flow step. Trained in 3.6 GPU-hours, SmolWAM succeeds in 95.5% of 400 LIBERO episodes at 57 ms and 2.3 TFLOP per 16-action chunk with 2.4 GiB of GPU memory, 11x”
Exact source · revision 1Source owner not reported
Reported quantity · FLOPs reduction: 18.5 other · Basis not reported
“faster than its teacher with 18.5x fewer FLOPs and 3.2x less memory, and its world head still forecasts scene motion.”
Exact source · revision 1Source owner not reported
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsAnolisVLA: Lightweight VLA for Mobile Manipulation with State Space Models
AnolisVLA improves task success and speed over SmolVLA and π₀.₅ in mobile manipulation. It uses state space models for efficient multimodal processing.
- arXiv Robotics — research abstractsSmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices
This paper introduces SmolWAM, a latent world-action model designed for robotic manipulation on edge devices. It is derived from the 2B-parameter Cosmos Policy by retaining its pretrained trunk and rebuilding the inference interface. SmolWAM achieves 95.5% success in LIBERO episodes with reduced computational and memory requirements compared to its teacher model.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).