NowWAM: Current Frame Denoising Outperforms Future Prediction in Robot Control
机器人学会“看现在”，比学会“想未来”更重要
A 2026 paper introduces NowWAM, a robot control method that trains diffusion models to denoise the current frame rather than predicting future ones. Experiments show comparable or superior success rates (up to 87.7% on LIBERO-Plus) while halving visual token processing and speeding up training by 1.8x. The study indicates benefits arise from exposing action modules to the full denoising trajectory, not prediction direction, working even with pure text-to-image models like Z-Image.
Source: Chinese robotics — Research news discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
82.89%
Reported trials
View original evidence
他们固定其他所有条件不变，只改训练时用的辅助目标是未来帧还是过去帧，结果发现：用未来16步和32步的画面当目标，成功率分别是82.89%和83.70%；换成用过去16步的画面当目标，成功率反而还高一点，83.80%。
Open source S13Reported success rate
83.7%
Reported trials
View original evidence
他们固定其他所有条件不变，只改训练时用的辅助目标是未来帧还是过去帧，结果发现：用未来16步和32步的画面当目标，成功率分别是82.89%和83.70%；换成用过去16步的画面当目标，成功率反而还高一点，83.80%。
Open source S13Reported success rate
79.1%
Reported trials
View original evidence
论文的消融实验给出了答案：保留两条视觉流（当前目标单独作为辅助目标）成功率是79.10%，比起当前帧去噪这种单流方案，明显要低一截，而且还得额外处理一倍的视觉信息量。
Open source S19Reported success rate
77.48%
Reported trials
View original evidence
消融实验显示，如果训练时只固定用"完全干净"的画面（也就是不引入任何噪声变化），成功率只有77.48%；而让噪声强度均匀分布，成功率能到83.46%；再把采样往低噪声方向偏移一点，成功率进一步提到84.97%。
Open source S21Reported success rate
83.46%
Reported trials
View original evidence
消融实验显示，如果训练时只固定用"完全干净"的画面（也就是不引入任何噪声变化），成功率只有77.48%；而让噪声强度均匀分布，成功率能到83.46%；再把采样往低噪声方向偏移一点，成功率进一步提到84.97%。
Open source S21Reported success rate
84.97%
Reported trials
View original evidence
消融实验显示，如果训练时只固定用"完全干净"的画面（也就是不引入任何噪声变化），成功率只有77.48%；而让噪声强度均匀分布，成功率能到83.46%；再把采样往低噪声方向偏移一点，成功率进一步提到84.97%。
Open source S21success rate difference
>7
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
他们固定其他所有条件不变，只改训练时用的辅助目标是未来帧还是过去帧，结果发现：用未来16步和32步的画面当目标，成功率分别是82.89%和83.70%；换成用过去16步的画面当目标，成功率反而还高一点，83.80%。
Open source S13
论文的消融实验给出了答案：保留两条视觉流（当前目标单独作为辅助目标）成功率是79.10%，比起当前帧去噪这种单流方案，明显要低一截，而且还得额外处理一倍的视觉信息量。
Open source S19
消融实验显示，如果训练时只固定用"完全干净"的画面（也就是不引入任何噪声变化），成功率只有77.48%；而让噪声强度均匀分布，成功率能到83.46%；再把采样往低噪声方向偏移一点，成功率进一步提到84.97%。
Open source S21
如果不这样做，而是只在最理想状态下训练，论文的数据已经摆在那了：77.48%对84.97%，差了超过7个百分点。
Open source S23
Source：Chinese robotics — Research news discovery · msn.cn