A research team from the Shenyang Institute of Automation, Chinese Academy of Sciences, together with Peking University and collaborating institutions, proposes the Embodied Context Protocol (ECP). Centered on semantic interfaces and declarative workflows, ECP is an interface protocol for orchestrating embodied systems, connecting simulation platforms, data acquisition, model training, and inference execution into a reusable and auditable workflow.

Background
Embodied intelligence is accelerating the shift from control driven robotics to model driven capabilities. In practice, however, real deployments still require engineering teams to integrate heterogeneous components such as robot middleware, simulation tools, model services, and industrial automation systems. Because these components often use different data conventions and coordination logic, systems typically rely on ad hoc scripts, state machines, and bespoke glue code. This makes workflows difficult to reuse, slow to redeploy, and costly to debug and maintain.

Research progress
The team presents the Embodied Context Protocol (ECP) as a layered interface protocol that specifies how task context is represented, how cross module interactions report progress and failures, how backend differences are adapted, and how workflows are composed for end to end orchestration. The paper also synthesizes recurring coordination bottlenecks in embodied deployments and abstracts key interface requirements, including context semantic consistency, capability declaration, task level composition, and consistent behavior across simulation and real backends.

What is ECP
ECP models an embodied task as context objects, executable interfaces, and traceable progress. It organizes interoperability into four layers:

• Semantic Layer. Transport agnostic schemas for observations, actions, and task context with explicit units, frames, and timestamps to reduce semantic drift.
• Interaction Layer. A bounded set of interface verbs plus uniform progress and failure envelopes to support supervision, timeouts, rollback, and recovery.
• Adapter Layer. Normalization and validation of units, frames, and clocks to support consistent execution across backends.
• Workflow Layer. Declarative composition of acquisition, training, inference, and execution into portable task graphs for reproducibility and auditing.

To make the protocol executable in system engineering, ECP maps the layered design to standardized interaction paths, organized as use cases, that connect simulation, data acquisition and storage, training and model management, inference, robot drivers, and industrial automation control into closed loop workflows.

Engineering example
The paper describes an implementable example in a pick and place task. Multimodal data are collected in a simulation environment to support training and workflow debugging. The resulting ACT (Action Chunking Transformer) policy is then deployed to an edge or fog server for inference. In execution, inference outputs are delivered through the robot control stack to the real system, while industrial process equipment is synchronized via PLC side signaling and status feedback. Across this chain, ECP runs through the full perception, inference, and control pipeline, effectively ensuring the consistency of context semantics and execution progress, and providing unified support for fault handling and traceable execution..

Future prospects

• For researchers. Reusable interface abstractions can reduce integration overhead and improve reproducibility when moving from prototypes to deployments.
• For engineering and industrial adoption. A unified orchestration layer can ease integration among models, simulators, robots, and industrial control systems, reducing redeployment cost across heterogeneous setups.
• For standardization. ECP is being advanced toward an industry standard for system interconnection. The team is promoting a proposal within China's electronics industry standardization process and aligning the work with the IES standard system. The roadmap also plans coordination with existing industrial automation and industrial internet standards for scalable humanoid robot and embodied production line applications.

The complete study is accessible via DOI:10.34133/research.1047