WHAT IT DOES

What does it do?

InstinctFlash is a high-performance serving framework for robot models. It detects the checkpoint’s model family, creates a manifest, and starts inference serving. It provides a Python API and a msgpack-over-WebSocket protocol so existing openpi clients can connect. A separate validation command checks whether a checkpoint can be published and records non-inferiority certification results in the package.

KEY FEATURES

Key features

• Inspect checkpoints to detect the model family and serving configuration
• Use isolated, fixed inference environments for each model
• Support native precision and optional FP8 precision
• Receive observation data through a Python API and return action arrays
• Serve over a network using msgpack-over-WebSocket
• Validate checkpoint publishability and non-inferiority