Robotics · 2026-10-02
FutureWorlds: Learning Robotic World Models from Alternative Futures
Hao Wu, Shengju Qian, Weiyan Wang, Fan Xu, Fan Zhang, Yuanpeng He, Qingsong Wen, Yuxuan Liang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
FutureWorlds learns robotic, action-conditioned world models by generating multiple candidate futures, maintaining candidate-specific histories, and learning from relative quality. It uses diverse beam search with bounded candidate memory and MemSPO, a GRPO-style update using group-relative video trajectory rewards. On RT-1, BridgeV2, and RoboCasa, the authors report LPIPS reductions for 32-frame predictions of 14.78%, 20.84%, and 9.12% versus the strongest baseline, with only 200 MemSPO updates further improving quality and enabling longer-horizon prediction.
Why it matters
The paper targets a key issue in learning from alternative futures: candidate comparisons are only useful if candidates are both diverse and history-consistent. By explicitly pairing each candidate with matching retained memory during both generation and policy scoring, and by converting rewards into group-relative advantages, FutureWorlds provides a coordinated way to turn alternative predictions into stable learning signals for robotic world models.
Method
- FutureWorlds uses a multimodal discrete autoregressive model for action-conditioned conditional video prediction and constructs alternatives with diverse beam search during RL training.
- Candidate-specific bounded persistent memory retains per-candidate scene/state anchors and rolling predictions, ensuring generation and policy scoring use matching histories.
- MemSPO converts decoded video trajectory rewards into group-relative advantages and applies a search-guided surrogate based on GRPO with clipped policy optimization.
Limitation
The paper notes that beam selection changes the candidate distribution, so MemSPO is a search-guided surrogate without an unbiased on-policy gradient guarantee.
Abstract (from arXiv)
Robotic world models predict action-conditioned future scenes, providing a foundation for understanding action outcomes. However, turning alternative predictions into useful learning signals remains challenging: similar candidates limit informative quality comparisons, while diverging trajectories require persistent maintenance of their individual histories. We introduce FutureWorlds, a framework that unifies candidate construction, history maintenance, and learning from relative quality. Built on a multimodal discrete autoregressive model, FutureWorlds uses diverse beam search during reinforcement learning to construct candidate futures that balance confidence and diversity. Candidate-specific bounded memory preserves scene states and ensures that generation and policy scoring use matching histories. We further propose MemSPO (Memory-Conditioned Search-Guided Policy Optimization), which converts video trajectory rewards into group-relative advantages to optimize the world model. On RT-1, BridgeV2, and RoboCasa, FutureWorlds reduces LPIPS for 32-frame predictions by 14.78%, 20.84%, and 9.12%, respectively, relative to the strongest baseline on each dataset. Under fixed evaluation configurations, only 200 MemSPO updates further improve generation quality and support continued prediction beyond the training horizon. Memory ablations, decoding sensitivity analysis, and optical-flow evaluation show that these gains extend beyond visual quality to more accurate motion prediction and more consistent object states. Project page and code: https://github.com/Alexander-wu/FutureWorlds.
Related papers
- One from Infinity: Actualizing Futures from Pretrained World Models into Robot Actions
- Forward Dynamics Models
- {\Delta}WAM: Distilling Action Tangent Fields into World Action Models
- MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling
- WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-01 · Indexed by Pub-AI: 2026-10-02 UTC
FutureWorlds: Learning Robotic World Models from Alternative Futures
Preprint posted: 2026-10-01 · Original record ↗
- arxiv
- 2610.01019
Explicit source versions (1)
- v1 · 2026-10-01
- Preprint available
Source date: 2026-10-01 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Robotic world models predict action-conditioned future scenes, providing a foundation for understanding action outcomes. However, turning alternative predictions into useful learning signals remains challenging: similar candidates limit informative quality comparisons, while diverging trajectories require persistent maintenance of their individual histories. We introduce FutureWorlds, a framework that unifies candidate construction, history maintenance, and learning from relative quality. Built on a multimodal discrete autoregressive model, FutureWorlds uses diverse beam search during reinforcement learning to construct candidate futures that balance confidence and diversity. Candidate-specific bounded memory preserves scene states and ensures that generation and policy scoring use matching histories. We further propose MemSPO (Memory-Conditioned Search-Guided Policy Optimization), which converts video trajectory rewards into group-relative advantages to optimize the world model. On RT-1, BridgeV2, and RoboCasa, FutureWorlds reduces LPIPS for 32-frame predictions by 14.78%, 20.84%, and 9.12%, respectively, relative to the strongest baseline on each dataset. Under fixed evaluation configurations, only 200 MemSPO updates further improve generation quality and support continued prediction beyond the training horizon. Memory ablations, decoding sensitivity analysis, and optical-flow evaluation show that these gains extend beyond visual quality to more accurate motion prediction and more consistent object states. Project page and code: https://github.com/Alexander-wu/FutureWorlds.