From Data Hours to Real Gains: Embodied AI's New Competition Paradigm

From Data Volume to Capability Gain

The embodied intelligence industry has long treated data scale — especially the "million-hour" benchmark — as a proxy for model capability. Companies raced to accumulate massive datasets, assuming more hours directly translated to better real-world performance. However, as robots move from lab demos to actual deployment, practitioners recognize that data duration and practical ability lack a simple exchange rate. The real world presents endless variations — changed object materials, shifted shelf positions, unexpected slippage — that training samples cannot exhaustively cover. Consequently, the evaluation criterion is shifting: instead of asking "how many hours," the field now asks "how much concrete capability gain does new data deliver?"

The Limits of Scaling Laws

Scaling laws do exist in embodied AI. At the 9th China Robot Summit, Xinghaitu's business director Shen Xin stated a production-grade model needs at least million-hour data accumulation. Zhejiang University professor Xiong Rong confirmed that model performance continues to improve with more training data. Yet, PoShell Robotics founder and Tsinghua assistant professor Xu Huazhe warned at the 2026 Bund Summit that a tenfold increase from 100k to 1M hours does not guarantee proportionally better results. He advocates "expand while searching, search while expanding" — first identify which data actually improves the model, then scale.

Defining Data Quality for Embodied AI

Ant Lingbo's chief data scientist Huang Yongtao separates quality into two tiers: an engineering baseline (no dropped frames, no distortion, multi-sensor clock sync) and a training-objective tier. For stable, smooth motions, consistent trajectories are valuable; for generalization, diverse object positions and shapes matter more. Physical Intelligence's π0 research adds a stage-based view: pre-training requires broad task and behavior coverage to build general capabilities, while post-training demands consistent, fluent demonstration policies for specific tasks. The research also shows that high-quality demonstrations alone lack failure samples needed for error recovery, while diverse but lower-quality data fails to yield efficient, robust policies. Thus, diverse experience and high-quality demonstration must complement each other across training stages.

This reframes data quality: a segment must be accurate and complete to enter training, but its value depends on whether it covers the model's current capability gaps and improves task performance. Data producers must therefore understand model needs deeply — deciding which experiences to expand, which samples to reduce, and which scenarios remain missing — turning data production from one-off delivery into a continuous process synchronized with model iteration.

Real2Sim2Real Closed Loop

Liu Shengxiang, founder of Wenwen Intelligence (无问智科), argues scale and quality are not inherently opposed. The key is a high-fidelity Real2Sim2Real loop: real-world data provides physical interaction grounding; simulation and generated data expand scene and task distributions; training, evaluation, and real-robot feedback mutually correct each other, keeping data production focused on the model's capability gaps. World models play a pivotal role, learning spatial structure, object relations, motion processes, and interaction laws from real data to understand, reconstruct, and generate reality, converting limited real experience into vast virtual training environments.

无问智科's Three-Stage Pipeline

Step 1: Real-World Collection with "无垠" Data Platform

Wenwen Intelligence uses proprietary multimodal devices for "unobtrusive wild collection," capturing human behavior, object interactions, environment, and physics during actual tasks. Unlike fixed collection sites with scripted actions, this preserves real-world scene variations and operational experience. Raw data undergoes cleaning, alignment, filtering, reconstruction, annotation, and cross-embodiment processing to become Model-Ready data. The company reports that its automated data production pipeline, powered by world understanding models, boosted Raw Data to Model-Ready Data processing efficiency by 500% to 1000%. They have deployed across logistics, manufacturing, home service, and retail, accumulating hundreds of thousands of hours of real physical interaction data with a monthly production capacity exceeding 20,000 hours.

Step 2: Generating Training Worlds with "无臻" Platform

Leveraging world models, the "无臻" platform understands, reconstructs, and generates real scenes, automatically producing Sim-Ready 4D assets and reinforcement learning environments. The goal is not mere replication but varying objects, layouts, task conditions, and interaction processes so models encounter more situations that could occur in reality. A single real scene can be structured and reconstructed into numerous new object configurations, action combinations, and task conditions, vastly expanding training coverage. Asset construction time dropped from 3 days per asset to 5 minutes, with a million-level Sim-Ready asset library built.

Step 3: Simulation and Validation with "无穹" World Simulator

The "无穹" simulator provides high-precision differentiable physics simulation for large-scale physical AI trial-and-error. Virtual environments enable parallel training processes while avoiding real-robot wear and safety risks. However, virtual training's value hinges on sim-to-real transfer. Sensor noise, contact errors, and object deformation create gaps. Therefore, Wenwen Intelligence does not stop at Real2Sim but closes the loop with Real2Sim2Real: training results are tested on real robots, and the feedback guides the next round of data collection, world generation, and simulation training.

Commercial Validation and Market Signals

The closed loop's practical value is verified in customer projects. In a collaboration with a leading client, online reinforcement learning in generated worlds raised simulation success rates from 9.7% to 79.8%; transferred to real robots, task success rates improved over 50%. Commercial traction confirms demand: Wenwen Intelligence reports embodied AI business orders reaching hundreds of millions RMB, with year-over-year order growth exceeding 3000% and projected revenue growth over 1000%. Headline client coverage and repurchase rates both exceed 80%. The company partners with multiple embodied AI firms and foundation model companies on real data, simulation, evaluation, and Real2Sim2Real loops. In September 2024, it closed a Series A round of hundreds of millions RMB led by Hongtai Fund, with follow-on from HongShan Capital, China Electronics Data, Lion City Capital, and oversubscribed participation from existing shareholders. Funds target world model R&D, real physical interaction data scaling, Sim-Ready asset systems, world simulator construction, and full-chain Real2Sim2Real capability enhancement.

Conclusion: Toward Verifiable, Transferable Capabilities

Order growth, client retention, and investment direction point to a clear trend: as robots enter factories, warehouses, and homes, data needs extend from "obtaining a batch of training data" to plugging model gaps, expanding training scenarios, lowering trial costs, and driving model iteration through continuous feedback. Data infrastructure value is shifting from delivery capability to training-loop capability. While data duration remains a key supply metric, the industry will increasingly judge by a scorecard closer to real results: which new tasks the robot masters, whether it adapts to environmental changes in real time, and how much virtually learned capability transfers to reality. Embodied AI's data competition will ultimately hinge on whether data converts into verifiable, transferable real-world capabilities — only those that survive real-task scrutiny can push robots into production sites.