Embracing the bubble = Embracing controversy
The golden window for investing in crypto is precisely those years of greatest controversy.
When it was still viewed as a "speculative tool" and "regulatory gray area," most people saw risk, while a few saw a new network yet to be priced by the mainstream.
Similar stories are not uncommon:
▌When the iPhone was first released, many believed that a phone could only be used for communication, and a small screen could never replace a PC.
▌Electric vehicles have long been questioned for their short range, high costs, and difficulty in recharging, and Tesla was at one point seen as a capital game.
The greater the controversy, the greater the cognitive disparity. By the time consensus forms and narratives are proven correct, the asymmetrical opportunities are often nearly gone.
1/ The next controversy: Robotics
One of the most contentious industries today is embodied intelligence/robotics.
Optimists believe it will trigger the next labor revolution; pessimists argue that it is merely an industrial version of bike-sharing wealth created through PowerPoint.
The skepticism is not unfounded:
High hardware and maintenance costs
Human-like forms may not be the optimal engineering solution
High-quality real-world data is extremely scarce
Application scenarios are dispersed, and willingness to pay is limited
These issues all point to the same core: friction in the physical world.
Language models can learn text and images from the internet, but robots must learn how to act.
They must not only recognize a cup but also determine where to grab, how much force to use, how to avoid obstacles, and how to set it down securely. If they miss, slip, or misjudge the position of an object, they must also know how to correct it.
This type of data is much scarcer than text on the internet.
2/ What robots lack is not stories, but data
The embodied intelligence industry is stuck in a clear contradiction.
On one side, hardware, sensors, computing power, and models are advancing rapidly;
on the other side, the data needed for robots to generalize in the real world remains expensive, slow, and scarce.
Traditional data collection relies heavily on closed laboratories: setting up scenarios → arranging operators → remotely controlling real machines → engineers cleaning, labeling, and training data.
The problems with this model are apparent:
▌Slow speed, high costs
▌Limited scenario coverage
▌Difficulty in handling the vast number of long-tail situations in the real world
▌Data supply struggles to keep up with model iteration speeds
So the real bottleneck in the industry may never have been about whether robots could be made, but whether data that makes robots smarter can be continuously produced at a sufficiently low cost and fast speed?
This is precisely the entry point of @axisrobotics.
3/ Axis Robotics: Not making robots, but producing robot data
Axis Robotics has not focused on creating another humanoid robot but has chosen a more fundamental and harder-to-scale layer in the industrial chain: robot training data.
Its approach is divided into three steps.
Step one: Use a browser to turn simulation data into a scalable business
Users do not need any hardware; they can remotely operate a simulated robotic arm via a browser to perform tasks such as picking up cups, sorting items, and opening drawers.
Every operation leaves behind a complete trajectory—joint states, object poses, control actions, task metadata, etc.—which can be directly used for pre-training basic strategies.
Step two: Use human correction to collect high-value post-training data
When the strategy begins to autonomously execute tasks, humans only need to supervise. If the model makes a mistake, the operator immediately takes over, corrects it, and then hands control back to the model.
Step three: Use smartphones to fill in the real-world gap
Smartphones can capture first-person spatial and motion information, directly translating everyday human behaviors into samples that robots can learn from, filling the gap that simulations struggle to cover.
The backend will replay, verify, and clean the trajectories, and then expand them into large-scale, photo-realistic, domain-randomized training samples using Isaac Sim, for training more generalized imitation learning strategies.
4/ Conclusion
Whether embodied intelligence can land as quickly as the market expects remains unknown. But one thing is becoming increasingly clear:
If physical AI truly enters the real world, data will be an indispensable infrastructure.
Axis Robotics transmission - https://s.kaito.ai/8QiHxmm
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
