Darren LI

AI研究的历史表明,相对简单、能够随着数据规模扩大的算法总是胜过那些复杂却无法随数据规模扩大的算法

基于物理引擎的仿真器,如Pybullet、Mujoco、Isaac Gym等,为获取大量机器人交互数据提供了一个有效的方式。研究者们可以先在仿真器中进行训练,之后再迁移到真实机器人上。

Key unlock: LLMs will be able to take into account vast amounts of relevant information and offer more personalized, tailored, and useful outputs.

第一,contextual data,上下文数据,AI对于语境的理解;第二,personal data,个性化数据,AI对于使用者的理解;第三,collaborative data,合作化数据,AI应用应处在工作流和交互中,而不是作为工具独立存在。

人形机器人每个关节都需要大电机,导致整体功率大、续航短。这里面需要新的结构设计、算法设计

3. I think this will most likely be the last cycle of cryptocurrencies. Because last cycle Bitcoin went from about 1200 to 20 000, it increased about 20 times, and this cycle from 20 000 to 60 000, it increased 3 times. The increase is much smaller than in the previous cycle. I think it is very likely that cryptocurrencies will trade like Nasdaq in... See more

What about larger companies? With some caveats, larger companies aren’t worried about going out of business because their Amplitude is too expensive. They might be price-conscious in terms of negotiating a contract in context of their budget, but most SaaS software is just a line item at the end of the day. What matters more is that the solution is... See more

对于locomotion,外界环境大多可以视为一个刚体,物理特性基本可以忽略。这使得可以花费更多时间来建立机器人本体的精确物理模型,以及设计更复杂的物理引擎。这是为什么RL更适合Locomotion