Configuration Reference

Configuration Reference

中文版

Configuration sources and precedence

Entry point	Configuration load order
run_web_client.py	conf/client_conf.py defaults → --conf YAML file (default: conf/default_conf.yaml) → runtime load/patch requests from the Web console
run_server.py	conf/server_conf.py defaults → command-line overrides (--model_type, --model_path, --debug)

The Web client accepts --host, --port, --conf, and --reload. The YAML file selected by --conf is both the startup configuration and the save target used by the Web console (Setup → Apply / Save As). When saving, .yaml/.yml targets are written as YAML and any other extension as a Python dictionary file.

Only existing leaf keys are updated when a configuration file or patch is applied; unknown keys are ignored. The default file conf/default_conf.yaml is git-ignored: it does not exist in a fresh clone and is created by the first save from the Web console; on startup it is applied only if present. Two legacy key names are migrated automatically on load: inter_chunk.search_length is read as inter_chunk.search_action.search_length, and robots.a2d.* is read as robots.agibot_g1.*.

All durations in the following tables are milliseconds unless another unit is shown explicitly. The panels that expose these parameters in the browser are described in Web Console.

Client root dictionary

get_client_config() contains these top-level sections:

Key	Description
rdm	Real-time observation buffering and inference-loop mode.
intra_chunk	Processing of actions within one model-generated action chunk.
inter_chunk	Fusion of the previous and next action chunks.
controller	Robot-command timing and execution speed.
robots	Selected robot adapter and adapter-specific settings.
vla_zmq	VLA inference-server connection and heartbeat settings.
visualize	WebSocket visualization server and display defaults.
record	Demonstration, evaluation, and LeRobot dataset recording.
vision	Observation-frame selection and preprocessing.
language	Language-task file and automatic subtask switching.

rdm: real-time data manager

Key	Type / valid values	Meaning
mode	async or sync	Inference/control loop mode. async avoids blocking on every inference result; sync waits for synchronized processing.
max_len	positive integer	Maximum number of observation frames retained in the buffer.
observe_fps_window_size	positive integer	Number of most recent captured frames used when estimating observation FPS.

intra_chunk: action processing inside a chunk

Key	Type / valid values	Meaning
intra_chunk_mode	raw, interpolation, or fitting	Action output mode: use raw actions, interpolate them, or use polynomial fitting.
filter_window_size	non-negative integer	Half-window used to filter gripper actions; the effective window is $2n+1$.
min_gripper_action_threshold	number	Gripper values below this threshold are treated as fully open (0.0).
max_gripper_action_threshold	number	Gripper values above this threshold are treated as fully closed (1.0).
fitting_num_samples	positive integer	Reserved trajectory-sampling setting; the current implementation uses the controller-period time grid instead.
fitting_deg	non-negative integer	Polynomial degree for fitted trajectories.
max_joint_fitting_workers	positive integer	Reserved worker setting; the current batch implementation does not create a joint-fitting executor.
max_gripper_fitting_workers	positive integer	Reserved worker setting; the current batch implementation does not create a gripper-fitting executor.
max_head_fitting_workers	positive integer	Reserved worker setting; the current batch implementation does not create a head-fitting executor.
joint_dim	non-negative integer	Number of dual-arm action dimensions processed by this stage.
gripper_dim	non-negative integer	Number of gripper action dimensions processed by this stage.
head_dim	non-negative integer	Number of head action dimensions processed by this stage.

These dimensions must match the selected robot’s action layout and the model output convention.

inter_chunk: fusion between chunks

Key	Type / valid values	Meaning
inter_chunk_mode	search_action, smooth_velocity, min_jerk, or sync	Method used to join consecutive action chunks.
search_action.search_length	positive integer	Number of future actions inspected while choosing a smooth join. Increase it if the robot visibly pauses at chunk boundaries.
smooth_velocity.max_vel	positive number	Velocity clamp used by velocity-smoothing fusion.
smooth_velocity.max_acc	positive number	Acceleration clamp used by velocity-smoothing fusion.
smooth_velocity.k_p	number	Proportional feedback gain applied to the position error.
smooth_velocity.k_d	number	Derivative feedback gain applied to the joint velocity.
smooth_velocity.kf_vel	number	Feedforward gain applied to the target velocity.
smooth_velocity.kf_acc	number	Feedforward gain applied to the target acceleration. Setting both feedforward gains to 0 degrades the fuser to PD feedback control.
min_jerk.blend_threshold	number	Threshold controlling when minimum-jerk blending is applied.
min_jerk.adaptive_factor	number	Adaptive scaling factor for minimum-jerk blending. A negative value keeps the implementation’s automatic behavior.
sync	mapping	Reserved configuration section for sync mode; currently empty. In sync mode the next action chunk is used directly without blending, and the Web console mode dropdown lists only search_action, smooth_velocity, and min_jerk.

controller: robot command timing

Key	Type / unit	Meaning
wait_time	number, ms	Delay before the next inference/control iteration. Increase it when command generation causes visible hesitation.
period	number, ms	Robot command period used by the control timer and trajectory timing.
speed	positive number	Execution speed relative to teleoperation speed.
raw_fps	positive number, Hz	Dataset FPS used to train the VLA/WAM model; used as the timing baseline.
gripper_offset	integer, frames	Forward offset for stepwise gripper commands relative to arm commands. Increase a positive value if the gripper reacts too late.

vision: observation preprocessing

Key	Type / valid values	Meaning
history_frame	boolean	Include a historical observation frame in addition to the current frame when the model input supports it.
preprocess.method	string	Name of the preprocessing function resolved from client.utils.misc; none disables preprocessing.
preprocess.keep_ratio	boolean	Preserve aspect ratio while resizing.
preprocess.height	positive integer, pixels	Target image height for preprocessing.
preprocess.width	positive integer, pixels	Target image width for preprocessing.

language: task instruction selection

Key	Type	Meaning
file_path	path relative to conf/	JSON task-instruction file, normally language_cmd.json.
task_id	string	Task key in the language JSON file. If absent, the first available task is used.
sub_task_id	zero-based integer	Index of the active instruction in the selected task. Invalid indices fall back to 0.
auto_mode_start_sub_task_id	zero-based integer	Sub-task index applied when Auto Mode starts. The Web console exposes it as the Start selector of the Language Command panel.
auto_mode	boolean	Automatically move to the next language instruction according to task-progress output.
task_progress_threshold	number	Mean progress probability required before automatic switching.
task_progress_win_size	positive integer	Sliding-window length used to average task progress.

language_cmd.json is a mapping from task IDs to ordered arrays of instruction strings. Add a task by adding one such array; do not store configuration metadata in this file.

vla_zmq: VLA server connection

Key	Type / unit	Meaning
ip	host name or IP address	VLA server address.
port	string or integer	VLA server TCP port; together with ip forms tcp://ip:port.
infer_timeout	positive integer, ms	Client receive timeout for one inference response. A timeout may drop a late response.
heartbeat_interval	positive integer, ms	Client heartbeat period.
heartbeat_timeout	positive integer, ms	Server-side inactivity time before a client is removed.

visualize: visualization service

Key	Type / unit	Meaning
host	host/IP	Interface on which the visualization WebSocket service listens.
port	integer	Visualization service port.
max_size	integer, bytes	Largest accepted WebSocket message.
ping_interval	number, seconds	WebSocket ping interval.
ping_timeout	number, seconds	Time allowed for a ping response.
updata_fps	positive number, Hz	Visualization update rate. The key name is retained for compatibility.
camera.open_head	boolean	Show the head camera.
camera.open_wrist_left	boolean	Show the left wrist camera.
camera.open_wrist_right	boolean	Show the right wrist camera.
trajectory.play	boolean	Start trajectory playback automatically in the UI.
trajectory.source	string array	Trajectory series displayed in the UI, such as State, ActionRaw, and ActionFitted.
trajectory.selected_joints	integer array	Initially selected action/state dimensions.
trajectory.window_span_sec	positive number, seconds	Time range displayed by the trajectory plot.

record: recording and LeRobot metadata

Key	Type	Meaning
switch	boolean	Master recording switch. It is read from the client configuration; the current run_web_client.py entry point has no --record option.
auto	boolean	Enable automatic recording behavior.
save_dir	project-relative path	Root directory for recorded data.
is_record_episode	runtime boolean	Whether episode recording is active.
is_record_eval_log	runtime boolean	Whether evaluation-log recording is active.
is_record_expe_data	runtime boolean	Whether experiment-data recording is active.
save_raw	boolean	Save source-resolution images rather than resized images.
evaluation.scores	numeric array	Score choices offered by the evaluation UI.
lerobot.codebase_version	string	Dataset/codebase version recorded in metadata.
lerobot.robot_type	string	Robot type recorded in metadata.
lerobot.total_episodes, total_frames, total_tasks, total_videos, total_chunks	non-negative integer	Dataset counters maintained in LeRobot metadata.
lerobot.chunks_size	positive integer	Number of episodes/frames represented by a storage chunk according to the recorder format.
lerobot.fps	positive number, Hz	Dataset frame rate.
lerobot.state_shape, action_shape	positive integer	Flattened state and action dimensions written to the dataset.
lerobot.splits	mapping	Dataset split definitions, for example {valid: "0:100"}.
lerobot.data_path	template string	Parquet path template relative to the recording root.
lerobot.video_path	template string	Video path template relative to the recording root.
lerobot.cam.head, cam.hand_left, cam.hand_right	mapping	Per-camera feature definition. Each has shape.height, shape.width, shape.channel, plus encode.codec, encode.is_depth_map, and encode.has_audio.

Supported video codecs include mp4v, avc1, XVID, and MJPG. Camera shapes must describe the frames actually written by the selected robot pipeline.

robots: robot adapters

robots.type selects unitree_g1 (default), agibot_g1, mock, ti5_t170c, or navi_wa2. Configure only the subsection for the selected adapter unless preparing reusable profiles.

Shared action-layout convention

An action_layout section divides a flat action vector into named segments. Every segment has start (inclusive), end (exclusive), and policy.

Policy	Meaning
gradual	Continuous dimensions, normally arm joints; suitable for smoothing/fitting.
stepwise	Discrete or near-discrete dimensions, normally grippers/hands.
manual	Robot/Web-controlled dimensions that are not inferred by the model. Place these after model-controlled segments.

Where available, presets provides UI reset/manual targets. Default and Custom are general presets; AgiBot G1 grippers also provide Open and Close, its wheels provide Forward, Backward, Left, and Right, and NAVIAI-WA2 hands provide Open and Finger.

robots.agibot_g1

Key	Meaning
hand_type	End-effector mode: gripper, hand_as_gripper, or hand. It selects camera names and proprioception layout.
camera.color_order	Pixel order returned by retrieve_observation() (rgb).
camera.ref	Reference camera name.
camera.names	Mapping from canonical names (head, hand_left, hand_right) to robot camera identifiers.
camera.non_ref_fetch_workers	Number of persistent worker threads used to fetch non-reference cameras.
proprio_names	Ordered proprioception groups used to assemble robot state.
gripper_freq, head_freq	Command frequencies for gripper and head control.
manual_arm_interval	Time interval in seconds between manually generated arm targets.
action_layout	AgiBot G1 arm, gripper, head, waist, and wheel action ranges and UI presets.

robots.mock

Key	Meaning
camera.ref, camera.names	Reference camera and mapping to dataset observation keys.
action_layout	Simulated arm and gripper action ranges and presets.
manual_arm_interval	Manual arm command interval in seconds.
state_action_range	Ranges used to select state/action fields from the mock dataset.
dataset_path	LeRobot-style dataset root. It must contain matching Parquet data and video directories.

robots.ti5_t170c

Key	Meaning
hand_type	End-effector mode; China Mobile Lingxi uses hand. It selects camera names and proprioception layout.
camera.color_order	Pixel order returned by retrieve_observation() (rgb; the adapter converts camera frames).
camera.backend	Camera transport; remote_forwarder subscribes to a video forwarder.
camera.forwarder_uri	ZMQ URI of the video forwarder (default tcp://192.168.1.100:9100).
camera.ref, camera.names	Reference camera and mapping to forwarder topics (head/color, hand_left/color, hand_right/color).
camera.warmup_s	Seconds to wait for camera streams after connecting.
camera.depth_names	Depth topic mapping. Kept empty: depth frames are not consumed by VLA inference and are neither subscribed nor sent to the server.
proprio_names	Ordered proprioception groups (arm, hand) used to assemble robot state.
hand_freq, head_freq	Command frequencies for hand and head control; 1 means every control tick.
reset_robot_pos	Full robot reset target vector (arms at the reset pose, hands open). Kept aligned with the arm Default preset and the console reset position.
manual_arm_interval	Time interval in seconds between manually generated arm targets.
action_layout	Arm (gradual, presets Default = reset pose / Custom = previous homing pose), hand (stepwise, Default/Custom), and head (manual, NECK joints, Web-only) action ranges and UI presets.

robots.navi_wa2

Key	Meaning
tt	Servo/control period in seconds.
gain	Servo gain.
hand_type	End-effector mode; NAVIAI-WA2 uses hand. It selects proprioception layout and hand presets.
camera.color_order	Pixel order of incoming frames (bgr; cv_bridge uses bgr8).
camera.topic_dict	Compressed ROS image topics keyed by head, hand_left, and hand_right.
action_layout	Continuous arm and stepwise hand action ranges.
reset_position	Full reset pose vector; ordering must match the NAVIAI-WA2 driver.

robots.unitree_g1

Key	Meaning
camera.color_order	Pixel order of incoming frames (bgr; cv_bridge uses bgr8).
hand_type	End-effector mode; Unitree G1 uses hand. It selects camera names and proprioception layout.
reset_robot_pos	Full robot reset target vector (26 dimensions, all zeros).
action_layout	Arm (gradual, Default/Custom presets) and hand (stepwise, six-finger values, Default/Custom presets) action ranges.

Server and model dictionary

get_vla_server_config() provides zmq, max_infer_workers, max_decode_workers, and models.

Key	Meaning
zmq	Same VLA network fields as vla_zmq; the server binds its ROUTER socket using the configured port.
max_infer_workers	Maximum concurrent inference workers. Set to 1 to disable concurrent inference.
max_decode_workers	Maximum image-decoding workers.
image_pad_and_resize	Enable aspect-ratio-preserving padding and resizing of decoded images on the server.
image_target_height, image_target_width	Target image dimensions in pixels used when image_pad_and_resize is enabled.
models.type	Selected model: mock, act, gr00t_n1, gr00t_n1_5, gr00t_n1_6, rdt, smolvla, go1, pi0, pi05, tao, dm05, or wallx.

Model-specific keys are as follows:

Model section	Keys	Meaning
models.mock	model_path	Mock-model path placeholder.
models.act	model_path	ACT checkpoint path.
models.gr00t	model_path, embodiment_tag, data_config_key	GR00T checkpoint plus supported embodiment and data-configuration identifiers.
models.rdt	model_path, config_path, vision_encoder_name_or_path, lang_embd_path	RDT checkpoint, model configuration, vision encoder, and language embedding paths.
models.smolvla	model_path, root_path	SmolVLA checkpoint and associated data/root path.
models.go1	model_path, exp_path, data_stats_path	GO1 checkpoint, importable experiment path, and normalization statistics JSON.
models.pi0, models.pi05	config, model_path, is_hand	OpenPI model configuration/checkpoint; is_hand enables hand-action handling.
models.tao	model_path, embodiment_tag	TAO checkpoint and a TAO-registry embodiment identifier.
models.dm05	model_path, repo_path, action_mode, backend, vision_trt_engine_path, diffusion_steps, vision_attn_implementation, execute_horizon	DM0.5 checkpoint and repository paths, action mode, inference backend with an optional TensorRT vision engine, diffusion-step count, and actions per control cycle.
models.wallx	model_path, repo_path, embodiment_tag, device, flow_steps	WALL-X checkpoint and repository paths, embodiment identifier, inference device, and flow-matching step count.

Paths must be valid on the inference-server host. embodiment_tag and data_config_key are model-library identifiers, not arbitrary labels; use values registered by the matching model installation.

Visualization ZMQ configuration

get_vis_zmq_config() is used by the standalone visualization ZMQ service.

Key	Meaning
client_ip, client_port	Address used by a visualization client.
server_ip, server_port	Bind address used by the visualization server; * binds all interfaces.

Logging configuration

conf/logging_conf.py is programmatic logging configuration rather than a YAML parameter dictionary. It writes timestamped files below logs/, moves prior-day files into date directories, and removes date directories older than seven days. Console logging defaults to INFO; file logging defaults to DEBUG. Use setup_logging(log_filename, logger_name) when adding an entry point.