Robot I/O Modes

robot_io controls how the policy reads robot state and sends low-level commands. It is separate from:

• inference_backend, which controls ONNX / TensorRT policy inference.
• motion_backend, which controls the reference motion source.

For sim2sim with sim_env/base_sim.py, keep the policy on --robot-io zmq. For sim2real deployment, choose one of the three modes below.

Quick Choice

Mode	Processes	Use when
--robot-io inline	policy only	Preferred real-robot path when the policy runs on a machine with unitree_interface; avoids the extra ZMQ bridge hop.
--robot-io zmq + scripts/real_bridge.py	policy + Python DDS bridge	Use when you want the original split-process bridge based on unitree_sdk2py, or when unitree_interface is unavailable.
--robot-io zmq + scripts/real_bridge_cpp.py	policy + unitree_interface bridge	Use when you want the split-process ZMQ contract but prefer the unitree_interface robot binding in the bridge.

tip

If hardware tests show violent joint shaking, or high-dynamic / high-speed motions look much worse than expected, suspect unstable ZMQ I/O latency first. Try --robot-io inline before retuning the policy or gains.

Inline

Inline mode creates the Unitree robot object inside BasePolicy. The policy loop calls robot.read_low_state() and robot.write_low_command() directly. No real bridge process is started.

uv run sim2real/rl_policy/tracking.py \
  --robot-io inline \
  --policy-config checkpoints/lafan-aa/policy-ec592bb4_lafan_100style_student-5000.yaml

Use this path first for real deployment when latency jitter matters. Add --robot-interface <robot_network_interface> only when the robot network interface is not the default eth0.

ZMQ With real_bridge.py

This mode keeps the policy and robot bridge in separate processes. The bridge uses unitree_sdk2py, publishes low_state over ZMQ, and applies low_cmd from ZMQ to the robot.

Terminal 1:

uv run scripts/real_bridge.py

Terminal 2:

uv run sim2real/rl_policy/tracking.py \
  --policy-config checkpoints/lafan-aa/policy-ec592bb4_lafan_100style_student-5000.yaml

Add --interface <robot_network_interface> to the bridge command only when the robot network interface is not the default eth0.

ZMQ With real_bridge_cpp.py

This mode keeps the same ZMQ contract as real_bridge.py, but the bridge uses unitree_interface for robot I/O.

Terminal 1:

uv run scripts/real_bridge_cpp.py

Terminal 2:

uv run sim2real/rl_policy/tracking.py \
  --policy-config checkpoints/lafan-aa/policy-ec592bb4_lafan_100style_student-5000.yaml

Add --interface <robot_network_interface> to the bridge command only when the robot network interface is not the default eth0.

Rule Of Thumb

Use inline for the normal real-robot deploy path. Use ZMQ bridge mode when you need process isolation, want to debug the policy and robot bridge separately, or are running sim2sim with base_sim.py.