admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-24 10:16:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(audio buffers): add security crop to avoid audio buffer overfilling

Browse Source
This commit is contained in:
CarolinePascal
2025-05-15 13:18:01 +02:00
parent 59e8f4572c
commit ad01ef19f4
1 changed files with 16 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/scripts/lerobot_record.py
+16 -8
View File
@@ -327,12 +327,6 @@ def record_loop(
preprocessor.reset() preprocessor.reset()
postprocessor.reset() postprocessor.reset()
if dataset is not None and robot.name != "lekiwi":
for microphone_key, microphone in robot.microphones.items():
dataset.add_microphone_recording(microphone, microphone_key)
else:
async_microphones_start_recording(robot.microphones)
# Create a buffer for audio observations (shifting window of fixed size over audio samples) # Create a buffer for audio observations (shifting window of fixed size over audio samples)
audio_buffer = { audio_buffer = {
microphone_name: np.zeros( microphone_name: np.zeros(
@@ -341,6 +335,14 @@ def record_loop(
for microphone_name, microphone in robot.microphones.items() for microphone_name, microphone in robot.microphones.items()
} }
if (
dataset is not None and robot.name != "lekiwi"
): # For now, LeKiwi only supports frame audio recording (which may lead to audio chunks loss, extended post-processing, increased memory usage)
for microphone_key, microphone in robot.microphones.items():
dataset.add_microphone_recording(microphone, microphone_key)
else:
async_microphones_start_recording(robot.microphones)
timestamp = 0 timestamp = 0
start_episode_t = time.perf_counter() start_episode_t = time.perf_counter()
while timestamp < control_time_s: while timestamp < control_time_s:
@@ -364,10 +366,13 @@ def record_loop(
# Transform instantaneous audio samples into a buffer of fixed size # Transform instantaneous audio samples into a buffer of fixed size
buffered_observation_frame = copy(observation_frame) buffered_observation_frame = copy(observation_frame)
for name in audio_buffer: for name in audio_buffer:
buffer_size = audio_buffer[name].shape[0]
# Remove as many old audio samples as needed # Remove as many old audio samples as needed
audio_buffer[name] = audio_buffer[name][len(buffered_observation_frame[name]) :] audio_buffer[name] = audio_buffer[name][len(buffered_observation_frame[name]) :]
# Add new audio samples # Add new audio samples
audio_buffer[name] = np.vstack((audio_buffer[name], buffered_observation_frame[name])) audio_buffer[name] = np.vstack(
(audio_buffer[name], buffered_observation_frame[name][-buffer_size:])
)
# Add the audio buffer to the observation # Add the audio buffer to the observation
buffered_observation_frame[name] = audio_buffer[name] buffered_observation_frame[name] = audio_buffer[name]
@@ -427,7 +432,10 @@ def record_loop(
if display_data: if display_data:
log_rerun_data( log_rerun_data(
observation=obs_processed, action=action_values, compress_images=display_compressed_images observation=obs_processed,
action=action_values,
compress_images=display_compressed_images,
log_time=time.perf_counter() - start_episode_t,
) )
dt_s = time.perf_counter() - start_loop_t dt_s = time.perf_counter() - start_loop_t