CoScene Secures Nearly 100 Million Yuan in Funding, Compliantly Delivers Over 200 Million Yuan in Real Robot Data Orders Overseas | Linear Portfolio

Today, coScene announced the completion of a new funding round of nearly 100 million RMB. In September, coScene became the first to compliantly deliver over a thousand hours of Unitree real-robot data overseas, with a pass rate exceeding 99%.

coScene is the first physical AI data science company to simultaneously achieve nine-figure RMB real-robot data orders, data asset rights confirmation, and compliant delivery. By constructing diverse real-world environments and complex events, deploying full-body motion control platforms and highly dexterous end effectors, it transforms pre-deployment sites into reinforcement learning environments and creates robot data that balances diversity with reliable demonstrations.

Linear Capital was the lead investor in coScene's angel round, and has accompanied the company's growth ever since.

On November 9, coScene announced the completion of a new funding round of nearly 100 million RMB. This round was led by Yonggui Fund, with TIANJI, Changzhou Zhirong, and others participating. The investors span state-backed sci-tech investment institutions, leading embodied AI companies, and industrial capital. The funds will be primarily allocated to:

• Pre-deployment environments that provide diverse data and capability evaluation for robots
• Data understanding models specialized for high-precision pose tracking and multi-dimensional annotation
• Data infrastructure accelerating robot deployment across "collection-training-simulation-inference"

coScene is the first physical AI data science company to simultaneously achieve nine-figure RMB real-robot data orders, data asset rights confirmation, and compliant delivery. In September, coScene became the first to compliantly deliver over a thousand hours of Unitree real-robot data overseas, with a pass rate exceeding 99%. Team members hail from Apple, Google, Alibaba, Inceptio Technology, and hold computer science degrees from Ivy League universities and top Chinese institutions (Project 985), with years of focus on data governance research and engineering for autonomous driving and robotics. Learn more about what coScene is doing for robot data

A growing body of research shows that with appropriate mixed training and data ratios, Action Data/Trajectory Data complements the general visual-linguistic knowledge of LLMs or VLMs, endowing physical AI with more powerful zero-shot generalization capabilities. A more definitive trend in robot deployment has taken shape. So the truly important question becomes — how can physical AI maintain high generalization performance while also executing body control and end-effector manipulation more efficiently and accurately?

RoboICL research found that across three tasks — peg insertion, towel folding, and block bridge-building — adding real-robot demonstration data significantly improved average task progress scores for GPT-6 Astra-controlled machines. coScene's answer is to work backward from the endgame: construct diverse real-world environments and complex events, deploy full-body motion control platforms and highly dexterous end effectors, transform pre-deployment sites into reinforcement learning environments, and create robot data that balances diversity with reliable demonstrations. Across more than 200 scenarios and 600 task types in various pre-deployment sites, coScene creates infinitely combinable variations of props, robot starting positions, task sequences, lighting conditions, unexpected events, and more — generating robot data with complete contextual relationships, full-body coordination, and multi-segment continuity.

From left to right: the success rate of combined "grasp, sleeve, twist" action sequences is significantly lower than single actions, but improves markedly with continuous robot data; full-body coordinated robot data compared to vision-and-hand-only data for task success rate; robot data combined with human demonstration yields higher task success rates. This type of data has not only sparked purchasing interest from numerous clients, but also complements human demonstration data to generate better commercial returns.

coScene creates diverse data based on more complex Tasks and Episodics; traditional image labels are insufficient to provide effective learning signals for physical AI. coScene is committed to providing clients with high-precision pose tracking and multi-dimensional annotation data post-processing services, which not only improves training outcomes but also helps clients avoid wasted costs from inaccurate internal annotation.

In an Episodic designed by coScene, a rag must be found by the robot, which then responds to unexpected events in its path. For this type of video data annotation, semantic and object labels are only the most basic requirements; spatial relationships, proprioception, context, and other elements are typically also included. Automated annotation can achieve declining marginal costs through far greater precision and accuracy than manual annotation.

From left to right: NVIDIA Flora, microagi nanoannot-1, concept×Mecka, and SOTA models compete on task completion evaluation, action description/temporal segmentation, and operational detail recognition — the specialized model wins. The data processing model behind automated annotation can also feed back to optimize upstream data pipelines. Beyond basic quality checks for dropped frames, clock synchronization, and joint limits, the greater value-add lies in on-demand customization based on gap data, with key guidance provided during production.

Today, coScene's business model comes from an infrastructure that reinforces data feedback loops and flywheels, which can continuously generate high-value data in pre-deployment environments and sell it to physical AI companies. In the future, when physical AI crosses the pre-deployment threshold, coScene hopes to continue using this software to help industrial/scenario operators with robot deployment needs, enabling their robots to continuously evolve in real-world scenarios and improve economic returns.

Industrial/scenario operator partners can use coScene's tools at any time to improve robot performance. coScene can also help partners manage their data assets, creating additional revenue beyond core business operations. Because high-value real-robot data from vertical scenarios will never lack buyers. For industrial partners using locally deployed open-source models, this software can also identify gap data needed to improve robot reliability through simulation evaluation, and complete demand matching and targeted purchases in coScene's data marketplace.

Buyers and sellers can complete data demand matching through coScene's data infrastructure. In coScene's view, data — as a high-value byproduct — will continue to enhance the commercial and social value of coScene as a data infrastructure provider.