DreamZero (21)
- https://hjfy.top/arxiv/2602.15922 | https://dreamzero0.github.io | agibot + NVIDIA | Seonghyeon Ye + Joel Jang
与 cosmos policy 稍有不同, video backbone (Wan2.1，冻结) 用作编码器而单一 DiT（可学习）从 noise 中解码出 flow v. DiT 利用 kv cache 实现了历史观测.
demo 亮点是 agibot-G1 穿鞋带以及 unseen 地前进按电梯按钮. 附录中有 train-infer 的 attn mask 设计和 infer 加速等不错的资料.
EgoScale (22)
- ⭐️⭐️⭐️ https://hjfy.top/arxiv/2602.16710 : https://www.alphaxiv.org/abs/2602.16710 | https://research.nvidia.com/labs/gear/egoscale/ | NVIDIA | Ruijie Zhang + Linxin Fan
这个 demo 比较精彩，用 sharpa hand 实现了使用电动螺丝刀、试管吸液和双指拧瓶盖.
Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos
Motus: A Unified Latent Action World Model
- tri-mot (VLM 用于理解指令 + video gen model + action expert)
WorldVLA
SimpleVLA
Forcy policy
LIFT
- yi wang
VLA-JEPA
dit4dit yy硕推荐
Latent Policy Steering 光流
LAPA: 早期提出 FDM 和 IDM 模型
End-to-end training of deep visuomotor policies 四大神仙
UniVLA: 提出了一种两阶段的训练来更好地提取 Latent Action
关于 latent-action pretraining drifting 的问题值得看看.
https://yipko.com/posts/work/pi0.7/
- offline: Conservative q-learning for offline reinforcement learning. 惩罚 OOD 动作