Event ·Multi-source

RoboTTT：机器人学习测试时训练，扩展上下文至8000时间步

Stanford SVL与NVIDIA Robotics联合提出RoboTTT，基于测试时训练（TTT）方法，将机器人模型原生扩展到8000时间步的上下文（约5分钟肌肉记忆），推理成本恒定。相比以往机器人策略仅能处理数帧（<0.1秒），该方法将上下文长度提升3个数量级。在电路板组装任务中，RoboTTT通过一次人类视频演示即可实现零样本模仿学习。实验显示，8K上下文预训练性能比1K高出62%，且从128到8K时间步的闭合环路性能持续提升无饱和迹象。

Current read

斯坦福和英伟达搞了个新方法RoboTTT，能让机器人记住5分钟内的动作历史，一次看人演示就能模仿，还能自己纠错。上下文越长效果越好，8K比1K强62%。

11 sources63° AI heatLast updated 2026/7/15 20:55:33

Evidence

11 sources

Primary sourceFei-Fei Li

2026/7/15 20:55:33View source

Related source 1arXiv cs.AI

2026/7/16 17:59:06View source

Related source 2Jim Fan

2026/7/15 15:25:02View source

Related source 3vLLM

2026/7/15 19:02:05View source

Related source 4NVIDIA AI

2026/7/16 16:02:57View source

Related source 5ollama

2026/7/14 00:01:12View source

Related source 6Pandaily

2026/7/14 07:58:13View source

Related source 7Fireworks AI

2026/7/14 14:51:38View source

Related source 8Together AI

2026/7/15 14:19:12View source

Related source 9elvis

2026/7/15 14:48:02View source

Related source 10Decoder

2026/7/16 09:07:58View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story