[Feature] Create inference script without encode video for fast test policy on hardware

[whats2000]

What this does

This PR adds a new script, lerobot_infer.py, providing a fast, pure inference mode for running robot policies without video encoding or dataset recording overhead.

Key features:

• New CLI entry point: lerobot-infer
• Real-time policy inference loop
• Teleoperation support (optional)
• Required camera streaming for policies that depend on vision
• Keyboard controls (p/Space, t, r, Esc)
• Automatic temporary metadata directory creation/cleanup
• Optional visualization via --display_data=true

This improves the workflow for quickly testing robot policies on hardware.

---

How it was tested

The script was validated using a generic configuration with robot, teleop, and at least one camera stream, which is required for most vision-based policies.

Testers must use their own robot IDs, teleop IDs, ports, and camera setup.

Example test command (replace with your own devices):

uv run lerobot-infer \
    --robot.type=<ROBOT_TYPE> \
    --robot.port=<ROBOT_PORT> \
    --robot.id=<ROBOT_ID> \
    --robot.cameras="<YOUR_CAMERA_CONFIG>" \
    --policy.path=<POLICY_PATH> \
    --task=<TASK_NAME> \
    --teleop.type=<TELEOP_TYPE> \
    --teleop.port=<TELEOP_PORT> \
    --teleop.id=<TELEOP_ID>

Observed results:

• Start and stop inference works correctly
• Episode reset works correctly
• Manual teleop control works successfully
• Mode switching (policy ↔ teleop) works as intended

---

How to checkout & try? (for the reviewer)

Minimal example:

uv run lerobot-infer \
    --robot.type=<ROBOT_TYPE> \
    --robot.port=<ROBOT_PORT> \
    --robot.id=<ROBOT_ID> \
    --robot.cameras="<YOUR_CAMERA_CONFIG>" \
    --policy.path=<POLICY_PATH>

With teleop:

uv run lerobot-infer \
    --robot.type=<ROBOT_TYPE> \
    --robot.port=<ROBOT_PORT> \
    --robot.id=<ROBOT_ID> \
    --robot.cameras="<YOUR_CAMERA_CONFIG>" \
    --teleop.type=<TELEOP_TYPE> \
    --teleop.port=<TELEOP_PORT> \
    --teleop.id=<TELEOP_ID> \
    --policy.path=<POLICY_PATH>

Optional visualization:

--display_data=true

Keyboard controls:

• p / Space → start/stop inference
• t → toggle teleoperation
• r → reset episode
• Esc → exit

[whats2000]

Additional Notes

Current inference is using the record script, but this introduces several issues:

• Slow episode transitions — each episode waits for video construction, which adds unnecessary delay during testing.
• No immediate interruption — policies cannot be paused quickly in urgent situations (e.g., to prevent unsafe actions or hardware risks).
• No convenient recovery path — switching to manual control is not supported, making it harder to recover or reset the robot during experiments.

This PR addresses all of the above by introducing a lightweight inference-only script with instant start/stop control, keyboard-based safety interruption, and optional teleoperation for recovery.

[Copilot]

Pull Request Overview

This PR introduces a new fast inference script (lerobot_infer.py) designed for testing robot policies on hardware without the overhead of video encoding or dataset recording. The script provides a streamlined workflow with real-time keyboard controls for policy execution, teleoperation switching, and episode management.

Key Changes

• Adds lerobot-infer CLI command for pure inference mode without recording
• Implements keyboard-driven control flow (play/pause policy, toggle teleoperation, reset episodes, exit)
• Supports optional real-time visualization via rerun when --display_data=true

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.

File	Description
src/lerobot/scripts/lerobot_infer.py	New inference script with policy execution loop, keyboard listener, temporary dataset handling for policy metadata, and support for single/multi-teleop configurations
pyproject.toml	Adds CLI entry point lerobot-infer to project scripts

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.