Driving · 2026-10-09
MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling
Jie Chen, Ruofei Bai, Yuxin Cai, Yifeng Zhang, Chengyang He, Jun Li, Wei-Yun Yau, Guillaume Sartoretti
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
MiniWAM replaces dense native future-feature prediction in World Action Models with compact targets learned by PRISM (Predictive Representations via Inverse Spatiotemporal Modeling). PRISM compresses privileged current–future transitions using inverse-dynamics supervision plus feature reconstruction; its encoder is frozen, and MiniWAM learns to predict PRISM targets and actions from current observations. The authors report 65× fewer future tokens, up to 8× speedup in world-action training, and competitive results at 0.25B parameters.
Why it matters
The paper argues that world-action co-training does not require predicting native visual futures. By defining the world-modeling target as compact PRISM representations rather than high-dimensional native future features, MiniWAM aims to improve the performance–efficiency trade-off while maintaining strong control performance across multiple robot manipulation benchmarks and feature spaces (WAN2.1 VAE and DINOv3).
Method
- Learn PRISM targets: a spatiotemporal bottleneck encoder compresses privileged current–future visual features; Stage 1 trains with inverse-dynamics supervision and feature reconstruction; posterior mean targets are used for policy training.
- Train MiniWAM: freeze PRISM encoder; predict PRISM target and robot actions from current observation (no future inputs at deployment).
- Use controlled comparisons: replace native future-feature targets with PRISM targets under matched training setup and evaluate success and training-efficiency speedups.
Abstract (from arXiv)
World modeling has emerged as an effective co-training objective for robot policies, giving rise to World Action Models (WAMs) that jointly predict actions and future states. However, most WAMs predict future states in the native representation space of pretrained visual backbones, resulting in high-dimensional targets with substantial training cost. We introduce MiniWAM, which instead predicts compact future representations learned from privileged current-future transitions. To construct these targets, we propose Predictive Representations via Inverse Spatiotemporal Modeling (PRISM), which combines inverse-dynamics supervision with feature reconstruction to emphasize control-relevant transition information while preserving useful future-state information. With the learned PRISM encoder frozen, MiniWAM is trained to jointly predict the resulting targets and robot actions from current observations. With 65$\times$ fewer native future feature tokens, MiniWAM consistently outperforms native future-feature prediction with both DINOv3 and WAN2.1 VAE features, while achieving up to an 8$\times$ speedup in world-action training. At 0.25B parameters, MiniWAM is already competitive with substantially larger WAMs on LIBERO, LIBERO-Plus, and RoboTwin 2.0 simulation benchmarks. Representation analyses further show that PRISM contributes behavioral structure beyond feature reconstruction alone. These results demonstrate that effective world-action modeling does not require predicting native visual futures, and that compact predictive representations provide a strong and substantially more efficient target for policy learning. The project page is available at: https://j1dan.github.io/MiniWAM.
Related papers
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-09 UTC
MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling
Preprint posted: 2026-10-08 · Original record ↗
- arxiv
- 2610.12194
Explicit source versions (1)
- v1 · 2026-10-08
- Preprint available
Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World modeling has emerged as an effective co-training objective for robot policies, giving rise to World Action Models (WAMs) that jointly predict actions and future states. However, most WAMs predict future states in the native representation space of pretrained visual backbones, resulting in high-dimensional targets with substantial training cost. We introduce MiniWAM, which instead predicts compact future representations learned from privileged current-future transitions. To construct these targets, we propose Predictive Representations via Inverse Spatiotemporal Modeling (PRISM), which combines inverse-dynamics supervision with feature reconstruction to emphasize control-relevant transition information while preserving useful future-state information. With the learned PRISM encoder frozen, MiniWAM is trained to jointly predict the resulting targets and robot actions from current observations. With 65$\times$ fewer native future feature tokens, MiniWAM consistently outperforms native future-feature prediction with both DINOv3 and WAN2.1 VAE features, while achieving up to an 8$\times$ speedup in world-action training. At 0.25B parameters, MiniWAM is already competitive with substantially larger WAMs on LIBERO, LIBERO-Plus, and RoboTwin 2.0 simulation benchmarks. Representation analyses further show that PRISM contributes behavioral structure beyond feature reconstruction alone. These results demonstrate that effective world-action modeling does not require predicting native visual futures, and that compact predictive representations provide a strong and substantially more efficient target for policy learning. The project page is available at: https://j1dan.github.io/MiniWAM.