AI BriefingAsynchronous Inference Technology That Solves Robot Control Latency
Key point
By using asynchronous inference that separates action prediction from execution, robot control latency is reduced and task speed is doubled.
Details
Recent robot policy models (ACT, OpenVLA, etc.) tend to predict Action Chunks rather than single actions. The existing sequential inference approach required the robot to wait until the next action chunk was computed, causing execution delays and reduced responsiveness to environmental changes.
To solve this, an asynchronous inference approach that separates action Prediction and Execution has been introduced. The system is largely divided into two components.
- PolicyServer: Performs inference on high-performance hardware and computes the next action chunk.
- RobotClient: Stores received actions in a queue and executes them, not stopping even while waiting for a new chunk.
gRPC is used for communication between the two components, guaranteeing about 5x faster performance compared to REST API.
As a result of applying this approach, task completion speed improved by about 2x even with the same policy, and immediate replanning became possible upon object grasping failure, greatly improving control flexibility.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Stanford Researchers Propose EXPO-FT for Universal Post-Training in Robotics