机器人GPT-3时刻震动硅谷！0行代码，看一遍秒会，黄仁勋李飞飞参投
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Generalist AI unveiled GEN-1.5, a robotics foundation model that reportedly learns manipulation tasks from seconds-long demonstrations without task-specific training. Its broad one-shot adaptation suggests large-scale physical-interaction pretraining could make programming robots more like prompting language models, though results remain company-reported and limited to short, simple tasks.
- GEN-1.5 achieved a reported 59% average success rate across 10 tasks after one 3–12 second demonstration and zero gradient updates.
- With five minutes of demonstrations and 10 gradient steps, average success reportedly rose to 83%.
- The model exhibited untrained adaptations such as using a dustpan instead of a brush, recovering from obstructions, and switching to two-handed manipulation.
- Generalist AI attributes these capabilities to eight months of large-scale pretraining and a still-improving scaling curve; independent validation and complex long-horizon evaluations are still lacking.
Sources (1)
机器人GPT-3时刻震动硅谷！0行代码，看一遍秒会，黄仁勋李飞飞参投
TL;DR - Generalist AI unveiled GEN-1.5, a robotics foundation model that reportedly learns manipulation tasks from seconds-long demonstrations without task-specific training. Its broad one-shot adaptation suggests large-scale physical-interaction pretraining could make programming robots more like prompting language models, though results remain company-reported and limited to short, simple tasks.
- GEN-1.5 achieved a reported 59% average success rate across 10 tasks after one 3–12 second demonstration and zero gradient updates.
- With five minutes of demonstrations and 10 gradient steps, average success reportedly rose to 83%.
- The model exhibited untrained adaptations such as using a dustpan instead of a brush, recovering from obstructions, and switching to two-handed manipulation.
- Generalist AI attributes these capabilities to eight months of large-scale pretraining and a still-improving scaling curve; independent validation and complex long-horizon evaluations are still lacking.