GPT-6 Astra Can Design Robot Experiments Itself, Letting AI Train AI
When general-purpose language models can directly debug and train their own robot strategies in simulation, the valuation narrative of independent embodied-AI companies—‘physical AI beats digital AI’—starts to crack.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Robot researchers split into three camps over Astra's impact
At an IROS dinner, CMU assistant professor Shi Guan Ya conducted a survey of roughly 100 roboticists, asking how Astra affected their views. Responses splintered into three camps: one believed general-purpose models solve everything—‘robotics is over’—making further research pointless; a middle faction saw general-purpose and physics models as two separate layers, each occupying its ecological niche; a third, radical group insisted general-purpose models would never solve today's specific robotics problems. Chen Zhe estimates the ‘solve everything’ camp represents about a third of respondents—a notable split at a conference historically focused on classical robotics like control and tactile sensing.
— Chen ZheAstra's most disruptive capability: designing and testing its own experiments
Astra operates at three levels. First, it acts as a direct policy model outputting end-effector positions. Second, it supervises and corrects VLA model outputs—Yinhe General's data showed Astra alone succeeds 26% of the time, but when supervising π0.5's output, success rates reached 48%. Third, most striking: Astra autonomously builds simulation environments in Blender, generates objects and properties, then debugs motion-control strategies—completing policy modeling, training, and optimization end-to-end. This is ‘physical RSI’: AI training AI in the physical world.
— Chen ZheAstra's spatial breakthrough was built on thousands of Blender projects
The spatial-reasoning leap stemmed primarily from data, not new methods. In May, the team hired large numbers of Blender modelers for large-scale trajectory annotation—explaining Astra's jump in Blender generation capability. Such costs amortize across multiple uses: 3D assets, video, robot training. General-purpose companies have structural advantage here. The field suspects egocentric first-person video in pretraining, though unconfirmed; from scaling-law logic, Chen Zhe sees no reason to exclude it.
— Chen ZheEmbodied startups survive by becoming reusable interfaces for general models
If general-purpose models handle task decomposition and spatial reasoning, what's left for embodied startups? Chen Zhe advises: become the reliable ‘interface’ general models repeatedly call, while defending true closed-loop advantages—tactile and force feedback, domain constraints, hardware reliability, failure recovery, and accumulated industry data. He compares automotive supply chains: Tier-1 suppliers wedged between OEMs and component makers get squeezed from both sides.
— Chen ZheFei-Fei Li's World Labs sale shows the exit path for Physical AI startups
World Labs was acquired by AMD for $8.2 billion in AMD stock. Li's team had built spatial-reconstruction model Atlas and acquired simulation company SceniX. The price may not reflect pure business value, but signals what a $1 trillion company will spend on Physical AI strategy. It warns other founders: the best near-term exit is likely acquisition by a large tech company, since such buyers typically value only the first target. Chen Zhe emphasizes China lacks comparable acquisition pathways.
— Chen ZheUnitree's stock collapse belies the strengthening case for quality hardware
Unitree debuted at a peak exceeding 440 billion yuan but fell 60%, though still above its 600+ billion IPO valuation. Chen Zhe attributes volatility to short-term capital in the new-stock float; real valuation emerges after August 2026 when ChiNext lockups lift and institutional money reprices. On fundamentals: H1 2026 revenue grew 48% YoY, with humanoid output rising. His logic: Astra proves general models accelerate physical-AI deployment, which amplifies demand for reliable hardware. Unitree resembles Nvidia in autonomous driving—stronger upper-layer models drive greater hardware-ecosystem value.
— Chen ZheShowing a video now teaches robot skills faster than days of fine-tuning
In-Context Learning is the new capability Gen1.5 and S1 showed this quarter: without parameter retraining, a few-second-to-few-minute demonstration video lets models learn new tasks, replacing hours or days of teleoperation plus fine-tuning. Strikingly, this transfer works on out-of-distribution tasks never seen during training and can compose multiple learned atomic skills into longer sequences. The two companies differ on implementation: Generalist uses shorter trajectories and first-person egocentric input with joint data; Skill AI emphasizes third-person video up to ten minutes, suited for full-body skill composition.
— Chen ZheFCC regulations now check robot imports by origin of core components
In July, the US FCC expanded certification scope to robots over 2 kg with mobile sensing, networking, and control—including vacuums and lawn mowers—tallying foreign component origin at the core-part level. A US-brand robot with 65% or more parts from China or third countries faces rejection. Unitree's current models like G1, R1, and H2, already certified, face no immediate impact, but new designs will encounter friction. Chen Zhe reads this as tariff-driven policy unlikely to disrupt decades of Chinese supply-chain depth; long-term solutions will likely mirror automotive: moving production overseas.
— Chen ZheIn their own words · checked verbatim
One evening at a dinner, CMU assistant professor Shi Guan Ya conducted a survey with roughly 100 researchers present, asking how Astra was affecting them.
其实我们在有一天晚上的一个晚宴上面 CMU的助理教授石关亚老师 其实做了一个survey 他的survey就是问在场的可能100多个研究人员 Astro对大家的影响
Chen Zhe13:16
One faction believes foundational models solve everything—‘robotics is over’. If general models keep advancing, nothing's left worth researching. A middle faction sees general-purpose and physics models as two separate layers; each occupies its ecological niche. An extreme faction insists general models can never solve the problems we face in today's physical research.
但是争议比较大的两派 可能一派就是认为原模型solve everything 就是robotics is over 就是我们没有什么问题 还特别值得研究 如果通用模型可以不断的推进前沿 可能还有一个走中的派别是认为 通用模型跟物理模型 或者robotics问题是两层分层的问题 可以各自占据各自的生态位 那还有一派比较极端的是什么 还有一派极端的是认为 通用模型永远解决不了 我们今天物理学研究的很多问题
Chen Zhe14:17
All those problems we thought would stump even experienced practitioners just turn to dust before Astra. They're trivially simple to it—just pierced straight through.
今天在Astro面前都灰飞烟灭 就是这些我们认为 认知上一个真人司机 可能都容易犯错的问题 其实在Astro面前 都是很简单的问题 就直接打穿了
Chen Zhe44:52
They're focused on solving general intelligence. Everything else is derivative. The whole company, including the Bartis team, shares this consensus. It's the overarching company belief: you have to see the miracle yourself to believe it.
就是他们在解决通用智能问题 别的都是此问题 他们公司里的Bartis团队的人 也都是这么想 都是有这个共识的 我就是整个公司大的共识 没有见到神迹的人 就是不会信的
Chen Zhe53:55
After Astra emerged, the case for solid hardware companies is definitely strengthened. Everyone saw how general intelligence accelerates and gains capability—which actually reinforces the value of having robust embodied platforms and quality hardware.
我觉得判断肯定是增强了 在Astro横空出世过后 其实让大家看到了通用智能的这种加速 和通用智能的这个能力 其实是进一步加强了一个 一个solid的本体公司 或者硬件公司的价值
Chen Zhe57:56
Now there's a much simpler method: whether using first-person or third-person video, I can quickly demonstrate, and it learns the motions much faster. Project delivery speeds up significantly.
今天可能有了一个 更简单的方法 无论是用第一视角的视频 还是用第三视角视频 我就可以快速的展示 这样它这个动作 复制的速度 或者说项目落地的速度 就会大大的提升
Chen Zhe1:19:11
Figures
| World Labs acquisition price | $8.2 billion (stock) | 4:06 |
| Astra control success rate (Yinhe General, solo) | 26% | 19:26 |
| Astra control success rate (Yinhe General, supervising π0.5) | 48% | 20:26 |
| Unitree H1 2026 YoY revenue growth | 48% | 57:56 |
| Unitree H1 2026 revenue | 1.152 billion yuan | 1:01:00 |
| Unitree 2026 full-year revenue forecast | ~2.5 billion yuan | 1:01:00 |
| FCC robot weight threshold for certification | over 2 kg (with mobile sensing, networking, control) | 1:30:20 |
| FCC foreign core-part origin threshold for rejection | 65% or higher | 1:31:21 |
Glossary
- In-Context Learning (ICL)
- Learn new tasks from a few-second-to-few-minute demonstration video without retraining model parameters.
- Physical RSI
- AI autonomously builds simulation environments, trains its own strategies, completing the feedback loop—AI training AI in the physical world.
- VLA
- Embodied policy model that directly outputs motor-control signals for robots based on visual input.
- World Action Model (WAM)
- Action model using video generation as backbone, minimizing reliance on the language-understanding layer.
- Robot Dojo
- Open-source embodied-task benchmark used to compare success rates across different policy models.
- FCC
- US Federal Communications Commission; now certifies robots with mobile sensing, networking, and control.
How to listen
Practitioners in embodied AI and robotics startups, venture capitalists, and anyone wanting to understand how general-purpose models are reshaping the robotics landscape.
If primary-market speculation doesn't interest you, skip the 61–70 minute section on IPO regulations.