请问HN:数十亿美元的资金投入,但机器人技术有什么变化?
HN的朋友们——如果这是个无知的观点,请指教:
在过去两年里,我们见证了令人瞩目的机器人融资情况——例如,Figure的估值约为390亿美元,融资超过10亿美元;Skild AI融资约为14亿美元;Physical Intelligence融资数亿;还有像Wayve的机器人出租车融资约为15亿美元,当然也不能忘记马斯克的Optimus机器人。
这是一波疯狂的资本潮流,但从核心瓶颈的角度来看,自2016年至2020年间,实际上发生了什么变化?
我们听说过视觉模型、强化学习的进展、扩散策略、更好的模拟以及多模态的具身模型,但这些是否真的在大规模上突破了泛化、可靠的操作或真实的模拟与现实之间的差距?
一些问题:
1. 我们是否在朝着能够在混乱的真实环境中有效工作的通用策略迈出了实质性的一步?
2. “机器人基础模型”是否能像大型语言模型(LLMs)对自然语言处理(NLP)所做的那样,解决数据瓶颈?
3. 操作技术是否已经超越了渐进式的改进?
4. 人形机器人是技术上的飞跃,还是仅仅是吸引资本的叙事?
5. 有哪些真实的研究论文或基准显示出显著的进展?
我真心好奇我们是否处于一个技术拐点,还是将再次面临严峻的物理、数据或硬件问题。
查看原文
HN folks - school me if this is an uninformed take:<p>In the last 2 years we’ve seen eye watering robotics funding - e.g., Figure with a ~$39B valuation and $1B+ rounds, Skild AI raising ~$1.4B, Physical Intelligence raising hundreds of millions, and autonomous systems like Wayve’s ~$1.5B robotaxi funding, not to forget Musk with his Optimus bots.<p>That’s an insane capital wave, but from a core bottleneck POV, what’s actually changed since 2016-2020?<p>We’ve heard about vision models, RL advances, diffusion policies, better sim, and multimodal embodied models but have any of these really cracked generalization, reliable manipulation, or true sim2real at scale?<p>Some questions:<p>1. Are we meaningfully closer to generalist policies that work in messy, real environments?<p>2. Do “robot foundation models” solve the data bottleneck the way LLMs did for NLP?<p>3. Has manipulation gone beyond incremental improvements?<p>4. Are humanoids a technical leap or just a narrative that attracts capital?<p>5. What are the real research papers/benchmarks showing step-change progress?<p>Genuinely curious whether we are at a technological inflection point or are we going to hit hard physics/data/hardware problems again.