World Models / Research Direction

研究方向

Long-horizon robot world model 可以先按三个对象看：planner、verifier、backbone。

June 2026

Planner

  1. Planner-selected rollout audit。从同一批状态出发，记录 planner 生成的候选动作、world model 评分和真实执行结果，检查高分候选是否真的更容易成功。参考 TD-MPC、GRASP、Closing the Train-Test Gap、WEAVER。
  2. Planner-aware world model。把 planner 会选到的候选轨迹加入训练或微调，减少 data actions 和 planner actions 之间的 train-test gap。关键对照是 data actions、random actions、planner actions 三组上的评分校准和闭环成功率。
  3. Risk-aware planning。让 planner 同时看 value、uncertainty 和 event signal，在模型分歧变大、接触阶段切换、任务事件不确定时缩短 horizon 或触发 replan。

Verifier

  1. Ensemble verifier。用 model cluster / heads 的 rollout 分歧做信号，判断候选未来是否离开模型可信区域。参考 PETS、NEUBAY。
  2. Event verifier。把 imagined future 解码成任务事件，检查接触、移动、抓取、放置、阶段切换是否真的发生。参考 EV-WM。
  3. Reachability verifier。检查未来状态能不能由动作真实到达，避免 planner 选择“看起来好但动作不可达”的未来。参考 WAV。

Backbone

  1. Regime-aware transition。让 transition function 区分自由运动、接触前、接触中、离开接触后的动力学，减少接触和混合动力学被平均掉的问题。参考 PRISM-WM。
  2. Variable-horizon transition。让模型根据状态预测不同长度的 latent jump，直接测每个状态还能被信任 rollout 多远。参考 Variable-Length Latent World Models。
  3. World-action backbone。把 observation、history、action sequence 和 future state 放进同一个 backbone，让 model 同时服务 policy evaluation、policy improvement 和 test-time planning。参考 WEAVER、Cosmos Policy、DreamZero。

References

  1. World Models: A Comprehensive Survey, arXiv:2606.00133, 2026.
  2. World Models for Robot Learning: A Comprehensive Survey, arXiv:2605.00080, 2026.
  3. Koray Kavukcuoglu and Jeff Dean, Gemini 3.5: frontier intelligence with action, Google DeepMind / Google, 2026.
  4. NVIDIA: Pretrained to Imagine, Fine-Tuned to Act: The Rise of World Action Models, 2026.
  5. Nicklas Hansen, Hao Su, Xiaolong Wang. Temporal Difference Learning for Model Predictive Control, arXiv:2203.04955, 2022.
  6. Mingwei Li et al. Prismatic World Model: Learning Compositional Dynamics for Planning in Hybrid Systems, arXiv:2512.08411, 2025.
  7. Closing the Train-Test Gap in World Models for Gradient-Based Planning, arXiv:2512.09929, 2025.
  8. GRASP: Actions Speak Louder Than Frames, arXiv:2602.00475, 2026.
  9. Event-Verified World Model for Long-Horizon Robotic Manipulation, arXiv:2606.13053, 2026.
  10. World Action Verifier, arXiv:2604.01985, 2026.
  11. NEUBAY: Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism, arXiv:2512.04341, 2025.
  12. Beyond the Next Step: Variable-Length Latent World Models for Long-Horizon Planning, arXiv:2606.21775, 2026.
  13. WEAVER, Better, Faster, Longer: An Effective World Model for Robotic Manipulation, arXiv:2606.13672, 2026.
  14. PointWorld: A World Model with 3D Point Flow, arXiv:2601.03782, 2026.
  15. DreamZero: World Action Models for Zero-Shot Robotics, arXiv:2602.15922, 2026.
  16. Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models, PETS, arXiv:1805.12114, 2018.