Skip to content

Latest commit

 

History

History
47 lines (29 loc) · 2.97 KB

README.md

File metadata and controls

47 lines (29 loc) · 2.97 KB

Inference

This directory contains deployment and inference code for running trained OpenPi policies on real robots. Implementations include:

  • Temporal chunk-wise smoothing — smooth execution across action chunks (e.g. exponential decay, temporal ensembling [2]) to reduce jerk and improve tracking.
  • RTC (real-time chunking) — real-time execution of action-chunking flow policies [1]: the policy server streams chunks; the client executes with low latency and optional smoothing.

Details and script options (e.g. --rtc_mode, --smooth_method) are in the platform READMEs below.

Two robot platforms are supported:

Robot Stack Description
Agilex ROS Noetic, Piper arms, RealSense cameras Dual-arm bimanual manipulation (e.g. cloth folding). OpenPi client; optional temporal smoothing / ensembling. See agilex/README.md.
ARX ROS 2, ARX X5 arms Dual-arm manipulation (e.g. hang cloth). RTC and temporal chunk-wise smoothing (sync, async, ensembling). See arx/README.md.

Two-machine setup (policy server + robot)

Inference is split across two machines:

  1. GPU host (e.g. 4090) — Run the policy server so the trained model is served over the network.
  2. Robot / industrial PC — Run ROS, cameras, arms, and the inference client script that sends observations to the server and executes the returned actions.

On the GPU host, from the repo root:

uv run scripts/serve_policy.py policy:checkpoint --policy.config=<train_config_name> --policy.dir=<checkpoint_dir> [--port=8000]

Example (use the same config name and checkpoint path you used for training):

uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_libero --policy.dir=checkpoints/pi05_libero/my_experiment/20000

Default port is 8000. The server binds to 0.0.0.0 so the industrial PC can connect.

On the industrial PC, start the robot stack and inference client. In the inference script (e.g. Agilex), set --host to the GPU host’s IP address (so the client connects to the policy server). Use the same --port as on the server (default 8000). See agilex/README.md for the full sequence (Piper SDK setup → ROS → inference script with --host <gpu_host_ip>).

Important: Set the language prompt in the inference code to match the training config. For AWBC-trained models, use the same prompt format (including the positive-advantage form). See agilex/README.md#prompt-and-awbc and the stage advantage README.

  • Agilex: agilex/README.md — Piper SDK ref, setup, serve_policy + inference flow, prompt / AWBC.
  • ARX: arx/README.md — RTC and temporal chunk-wise smoothing scripts, device setup, inference flow, data collection.