DeepPolicy originated from Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences. With engineering-oriented reinforcement learning at its core, DeepPolicy provides autonomous optimization industrial agent systems for process industry, energy, semiconductor, and refrigeration sectors — advancing industrial control from reactive response to proactive optimization.
Research Origin
The team has spent years in reinforcement learning, optimal control, and industrial system modeling, undertaking national key research programs and publishing 40+ papers at top venues including NeurIPS, ICML, and AAAI — establishing a reliable engineering pathway from frontier algorithms to real industrial processes.
Continuous investment in safety constraints, sample efficiency, and offline training for engineering-deployable RL.
Distilling laboratory algorithms into deliverable platform foundations: ReinforceOS / ReinforceLab / ReinforceBox.
Deep engagement in petrochemical, energy, and data center frontlines — feeding real operating conditions back into research for closed-loop iteration.
Milestones
Key milestones at each stage, witnessing how algorithms progressively land in real industrial environments.
Deep research in reinforcement learning algorithms, successfully overcoming challenges in environmental disturbance and observation information trends. Published multiple papers in top international journals and obtained patent authorizations in China, the US, Europe, and Japan.
Industrial agent took initial shape — core technology completed real-world validation at a factory in Tokyo, Japan, earning industry recognition. Concurrently completed collaboration with Huawei on refrigeration system energy-saving.
Company officially established and operational, with the overall industrial agent architecture substantially refined. Broadly seeking compatible scenarios and partners across process industries while accelerating technology implementation and commercialization.
Industrial agent achieved scenario-specific application validation. Actual deployments began in petrochemical and material chemistry sectors for quality and efficiency gains; deployments completed in refrigeration systems and data centers to reduce cooling energy consumption.
Officially launched ReinforceOS, ReinforceLab, and ReinforceBox — three major products forming the complete 'Platform — Validation — Terminal' closed loop.
Ecosystem Partners
We have established long-term partnerships with leading DCS suppliers and process industry enterprises in China, completing algorithm validation and scaled deployment on real equipment.
Deep collaboration with mainstream DCS systems, delivering policies to process plants.
Algorithm validation completed on real operating conditions at petrochemical and chemical plants.
Cooling group control / PUE optimization / energy management validated in production.
Leveraging CAS research resources to continuously advance algorithm frontier research.
Deep partnerships established with 10+ leading enterprises, continuously expanding.
Core Capabilities
Building a complete capability stack across algorithm, platform, and terminal layers — making policies trainable, verifiable, deployable, and maintainable.
RL-based autonomous decision agents that continuously seek optima under constraints, replacing manual tuning and rule-based systems.
An integrated engineering platform for data ingestion, environment modeling, policy training, and online evaluation for industrial processes.
Algorithm verification, policy replay, and return estimation on offline data — reducing field trial-and-error costs.
Industrial-grade AI control box supporting DCS/PLC integration for low-latency closed-loop control on-site.
Cloud training / edge deployment / bidirectional sync — policies can be remotely upgraded and field-adaptive.
Continuous optimization around PUE, unit consumption, and stability metrics to deliver measurable business returns.
Constrained RL and safe exploration mechanisms to strictly protect process boundaries and production safety.
Templates covering petrochemical, power, water, semiconductor, and data center sectors for rapid scenario migration.
Industry Scenes
Across petrochemical, material chemistry, paper manufacturing, and refrigeration systems, each real-world site feeds operating conditions back into the algorithms, continuously refining the engineering boundaries of RL.
Industrial agents aggregate tens of millions of data points from plant zones, pipelines, and logistics in real time, dynamically optimizing reaction temperature, pressure, and feedstock ratios while ensuring safety interlocks, improving product quality consistency, reducing energy consumption and carbon emissions, and providing early warnings for leaks and corrosion.
Industrial agents analyze multi-variable coupling across batching, reaction, separation, purification, drying, and calcination processes to auto-optimize formulations and process parameters, reducing batch quality variance and shortening the cycle from pilot to mass production.
Industrial agents coordinate pulping, stock preparation, drying, and reeling processes — moving beyond fixed expert rules to adjust consistency, machine speed, and steam usage in real time, stabilizing paper basis weight and moisture while reducing cost waste and dryer energy consumption.
Industrial agents combine cooling load prediction with environmental changes to autonomously control compressor start/stop, frequency, condensing pressure setpoints, and air outlet volume, keeping refrigeration efficiency on the optimal COP curve, maintaining stable space temperatures, and delivering anomaly diagnostic recommendations.
Mission & Vision
We believe the deciding factor in next-generation industrial control is not more sensors or bigger databases — it's "self-optimizing strategy."
With RL at the core, deliver verifiable, continuously-evolving autonomous optimization agents to process industry, energy, semiconductor, and refrigeration sectors — helping real production shift from "reactive response" to "proactive optimization."
Build a full-chain product system covering platform, validation, and terminal — enabling every critical plant to run safer, more efficiently, and with lower carbon footprint through continuous algorithmic collaboration.
Contact Us
If your enterprise is looking for a deployable industrial intelligence optimization path, we welcome you to get in touch.
Research Center Shenzhen
Building F, 3F, 1068 Xueyuan Ave, Xili, Nanshan District, Shenzhen
SIAT, Chinese Academy of Sciences · Materials AI Research Center
R&D Center Beijing
Building 9, 4F, Room 406-2185, No.10 Automotive Museum West Rd, Fengtai, Beijing
Beijing DeepPolicy Technology Co., Ltd.
Business Center Ma'anshan
Ma'anshan Software Park E1, Huashan District, Ma'anshan, Anhui
Ma'anshan DeepPolicy Technology Co., Ltd.