I researched the data track of (robot) physics AI: I have previously seen an engineer from OpenAI working in a company in Shenzhen, mainly focusing on physical AI brains. Currently, they mainly pre train basic robot skills through open videos (or their own closed data). Then fine tune based on robot operation data in real environments, and even perform RL learning later. There is another type that mainly learns in the simulator environment and then further learns from the data transferred to the real environment. And the core here is the need for data and simulators that generate data. As the saying goes, as much intelligence as data is needed... And generating data, there are many types of companies in the market now, including crowdsourcing, factory cooperation, traditional giants (robot and data giants), and web3 on chain models that combine crowdsourcing, such as @ axisrobotics Web3 Chain Crowdsourcing Class: Axisrobotics @ axisrobotics is a web3 model crowdsourced physical AI data engine created by Princess Christine (✱, ✱) @ 0xsexybana. They mainly obtain data in two ways. One is to use a simulated operation robot simulator in the browser to obtain the skill behavior data of the robot. In order to obtain data in different scenarios at the same time, the simulator uses two simulators developed by Google and Nvidia, MuJoCo and IsaacSim, and has developed its own framework to integrate a single data into the cross simulator framework of the model pipeline. One is a user centric mobile application that achieves zero threshold operation. This application combines the most advanced real-time hand posture tracking technology, which can transform human vision into robot actions on a large scale - covering various workflows in daily activities and specific fields, and supporting modular configurations from single iPhone to multi camera headbands, wrist cameras, and UMI robotic arms. The final output result is 4D hand pose reconstruction, accompanied by rich language annotations. https://docs.axisrobotics.ai/ Factory cooperation crowdsourcing category: The representative of factory cooperation is Praxis Robotics @ praxisnation, whose vision is to "turn every company into a data supplier". Currently, he works closely with listed companies and unicorn level enterprise groups, serving 60000 employees across five continents and covering more than 150 environmental types. Collect daily behaviors of workers through proprietary hardware. https://www. ((praxisrobotics.io)) The other one has a more diverse model, supporting cooperation with factories as well as individual applications. http://tryasimov.ai If factory employees engage in physical labor, they hope to collect this data. In exchange, the wages of factory employees will be paid. For larger scale collaborations, support the provision of data sales revenue sharing. http://hub.xyz It is also a representative of this type, covering over 100 industries. It is necessary to apply online, verify and sign contracts for each store, and prepare for data collection to convert a large amount of physical work into data for training robots. Small and medium-sized enterprises participating in data collection will receive data sales and be divided into https://hub.xyz/ Cortex AI has also created a physical AI network called Cortex Marketplace, which supports individuals and small and medium-sized enterprises to apply for vertical industry and job behavior data. On this platform, real workplaces can earn rewards through self-centered robot motion capture and remote robot control sessions. https://cortexrobot.ai/ http://humanarchive.ai Using multimodal (head mounted camera+tactile gloves+IMU, etc.) real human data, focusing on India's gig economy With over 100 full-time team members, we collect large-scale multimodal video and sensor data from more than 100000 contributors and over 500 industry partners from households, hotels, restaurants, agriculture, manufacturing, construction, and retail industries in over 30 countries. There are many such enterprises, such as http://(senseirobotics. com), http://www. ((praxisrobotics.io))， http://www.luel.ai ， http://www.xdof.ai Domestic JD has also started using crowdsourcing to collect robot data in Suqian, China. JD plans to mobilize over 100000 local citizens (especially full-time mothers) to wear lightweight JoyEgoCam while doing daily household chores. I have previously written an introduction on x: https://(x.com)/alaching/status/2058725838179463465 Hardware for training data: Http://(robocap. io)) provides a 360 degree camera hat, which can be worn by any worker to collect 360 degree behavioral data during work. This company currently only sells this type of hardware and does not have a data crowdsourcing market. https://((robocap.io))/ Http://(ropedia. com) also provides a 360 degree helmet style collection hardware, but it is a full stack physical AI enterprise that provides data services, hardware, and trains its own models. Traditional robots and data giants (crowdsourcing): Robot http://figure.ai The company also provides a crowdsourcing platform, which has downloaded over 26000 videos and uploaded 16 million videos through a mobile app, covering 108 countries. They paid a compensation of 15 million US dollars for this. Interested parties can apply https://www.figure.ai/index-app . The data giant Scale has clearly not given up on this market. It mainly collects data from various industry work scenarios, including home, commercial, and industrial scenarios, through a global data collection network (self built), using different hardware such as two handed robotic arms, scale style seat belts, and customized hardware managed and operated by Scale. https://(scale.com)/physical-ai In the LLM era, the data industry has produced companies like Scale and Surge AI, which have billions of dollars in revenue and the best cash flow. However, they also appear to be fiercely competitive fields with various models. In the era of physical AI, the value of data remains important, and various models are flourishing.