srb agent teleop — Teleoperate Agent
Control a robot with an input device.
Usage
srb agent teleop --env ENV_ID [options]
For native deployment, srb real_agent teleop requires generated
COMMAND_MAPPING metadata and supports direct control with ROS, SpaceMouse or
haptic input. Its default device is ROS; keyboard, gamepad and policy-command
arguments are unavailable on that path. See the
native direct-input contract.
Options | Shared Agent Options
| Option | Description | Default |
|---|---|---|
--teleop_device [DEV …] | Interface (keyboard, spacemouse, ros, gamepad, haptic) | env default, else [keyboard] |
--pos_sensitivity VAL | Translation sensitivity | 1.0 |
--rot_sensitivity VAL | Rotation sensitivity (rad/s) | π (≈3.14) |
--invert_controls/--invert | In manipulation mode, flip the first two translational controls. Boolean optional forms --no-invert_controls and --no-invert disable inversion. | True |
| Teleoperation via Policy | ||
--algo ALGO | Algorithm of the policy (RL or IL) | |
--model MODEL | Path to the policy checkpoint or an SRB model artifact directory; this exact policy is loaded | |
--obs {state,visual} | Observation modality of the policy: selects the agent-config variant and the log tree searched for its latest checkpoint | state |
SpaceMouse read failures clear buffered motion and later valid callbacks can recover. Empty HID polls retain held input and count as fresh reads; buffered motion clears after 0.5 s without a successful read. A device that reads while reports are lost can hold its last motion; this is not a liveness check or hardware stop. Reset discards in-flight reads, close rejects late callbacks, and failed reads preserve the gripper toggle.
Teleoperation Modes
The action space determines direct or policy control:
Direct Teleoperation
Direct teleoperation is available for environments with action spaces that can be mapped into high-level control commands, such as the desired end-effector pose or acceleration of a mobile robot.
Examples
srb agent teleop -e _manipulation env.robot=ur10
srb agent teleop -e _orbital --teleop_device spacemouse
Teleoperation via Policy
For action spaces without direct controls (such as humanoid joint-space), use
a trained RL or imitation-learning policy. The device drives command-like
observations; the policy generates actions. Any algorithm accepted by
srb agent eval works.
Examples
Reference: Reinforcement Learning Workflow
Reference:srb agent train— Train Agent
Train a policy, then run teleop with it:
srb agent train --headless --algo dreamer -e locomotion_velocity_tracking env.num_envs=256
The latest checkpoint loads by default. Pass --model to select one:
srb agent teleop --algo dreamer -e locomotion_velocity_tracking --model logs/locomotion_velocity_tracking/dreamer/...
Public Demo Launch
excavation_demo uses one environment, manual reset, a long horizon, visible
HUD lines, and the high-quality RTX preset. Launch it with:
srb agent teleop --env excavation_demo
Add --hide_ui to drop the Isaac Sim UI for a fullscreen presentation station:
srb agent teleop --env excavation_demo --hide_ui
Task-configured devices (spacemouse, haptic, keyboard) make
--teleop_device optional. Hydra overrides such as env.scenery=lunalab and
env.demo_regolith_quality=visual remain available.