NERDDISCO's picture
fix: more types
ea0b085
import { WebTeleoperator } from "./teleoperators/base-teleoperator";
import { MotorConfig } from "./types/teleoperation";
import * as parquet from "parquet-wasm";
import * as arrow from "apache-arrow";
import JSZip from "jszip";
import getMetadataInfo from "./utils/record/metadataInfo";
import type { VideoInfo } from "./utils/record/metadataInfo";
import getStats from "./utils/record/stats";
import generateREADME from "./utils/record/generateREADME";
import { LeRobotHFUploader } from "./hf_uploader";
import { LeRobotS3Uploader } from "./s3_uploader";
// declare a type leRobot action that's basically an array of numbers
interface LeRobotAction {
[key: number]: number;
}
export class LeRobotEpisode {
// we assume that the frames are ordered
public frames: NonIndexedLeRobotDatasetRow[];
/**
* Optional start time of the episode
* If not set, defaults to the timestamp of the first frame
*/
private _startTime?: number;
/**
* Optional end time of the episode
* If not set, defaults to the timestamp of the last frame
*/
private _endTime?: number;
/**
* Creates a new LeRobotEpisode
*
* @param frames Optional array of frames to initialize the episode with
* @param startTime Optional explicit start time for the episode
* @param endTime Optional explicit end time for the episode
*/
constructor(
frames?: NonIndexedLeRobotDatasetRow[],
startTime?: number,
endTime?: number
) {
this.frames = frames || [];
this._startTime = startTime;
this._endTime = endTime;
}
/**
* Adds a new frame to the episode
* Ensures frames are always in chronological order
*
* @param frame The frame to add
* @throws Error if the frame's timestamp is before the last frame's timestamp
*/
add(frame: NonIndexedLeRobotDatasetRow) {
const lastFrame = this.frames.at(-1);
if (lastFrame && frame.timestamp < lastFrame.timestamp) {
throw new Error(
`Frame timestamp ${frame.timestamp} is before last frame timestamp ${lastFrame.timestamp}`
);
}
this.frames.push(frame);
}
/**
* Gets the start time of the episode
* If not explicitly set, returns the timestamp of the first frame
* If no frames exist, throws an error
*/
get startTime(): number {
if (this._startTime !== undefined) {
return this._startTime;
}
const firstFrame = this.frames.at(0);
if (!firstFrame) {
throw new Error("Cannot determine start time: no frames in episode");
}
return firstFrame.timestamp;
}
/**
* Sets an explicit start time for the episode
*/
set startTime(value: number) {
this._startTime = value;
}
/**
* Gets the end time of the episode
* If not explicitly set, returns the timestamp of the last frame
* If no frames exist, throws an error
*/
get endTime(): number {
if (this._endTime !== undefined) {
return this._endTime;
}
const lastFrame = this.frames.at(-1);
if (!lastFrame) {
throw new Error("Cannot determine end time: no frames in episode");
}
return lastFrame.timestamp;
}
/**
* Sets an explicit end time for the episode
*/
set endTime(value: number) {
this._endTime = value;
}
/**
* The time difference between the start and end time of the episode, in seconds
*/
get timespan() {
const hasNoFrames = this.frames.length === 0;
if (hasNoFrames) return 0;
return this.endTime - this.startTime;
}
/**
* The number of frames in the episode
*/
get length() {
return this.frames.length;
}
/**
* Creates a new LeRobotEpisode with frames interpolated at regular intervals
*
* @param fps The desired frames per second for the interpolated episode
* @param startIndex The desired starting index for the episode frames, useful when storing multiple episodes
* @returns A new LeRobotEpisode with interpolated frames
*/
getInterpolatedRegularEpisode(
fps: number,
startIndex: number = 0
): LeRobotEpisode {
if (this.frames.length === 0) {
return new LeRobotEpisode([], this._startTime, this._endTime);
}
const actualStartTime =
this._startTime !== undefined
? this._startTime
: this.frames[0].timestamp;
const actualEndTime =
this._endTime !== undefined
? this._endTime
: this.frames[this.frames.length - 1].timestamp;
const timeDifference = actualEndTime - actualStartTime;
const numFrames = Math.max(1, Math.floor(timeDifference * fps));
const interpolatedFrames: NonIndexedLeRobotDatasetRow[] = [];
const firstFrame = this.frames[0];
const lastFrame = this.frames[this.frames.length - 1];
for (let i = 0; i < numFrames; i++) {
const timestamp = actualStartTime + i / fps;
let frameToAdd: NonIndexedLeRobotDatasetRow;
if (timestamp < firstFrame.timestamp) {
frameToAdd = { ...firstFrame, timestamp };
frameToAdd.frame_index = i;
frameToAdd.index = startIndex + i;
} else if (timestamp > lastFrame.timestamp) {
frameToAdd = { ...lastFrame, timestamp };
frameToAdd.frame_index = i;
frameToAdd.index = startIndex + i;
} else {
let lowerIndex = 0;
for (let j = 0; j < this.frames.length - 1; j++) {
if (
this.frames[j].timestamp <= timestamp &&
this.frames[j + 1].timestamp > timestamp
) {
lowerIndex = j;
break;
}
}
const lowerFrame = this.frames[lowerIndex];
const upperFrame = this.frames[lowerIndex + 1];
frameToAdd = LeRobotEpisode.interpolateFrames(
lowerFrame,
upperFrame,
timestamp
);
frameToAdd.frame_index = i;
frameToAdd.episode_index = lowerFrame.episode_index;
frameToAdd.index = startIndex + i;
frameToAdd.task_index = lowerFrame.task_index;
}
interpolatedFrames.push(frameToAdd);
}
return new LeRobotEpisode(
interpolatedFrames,
actualStartTime,
actualEndTime
);
}
/**
* Interpolates between two frames to create a new frame at the specified timestamp
*
* @param frame1 The first frame
* @param frame2 The second frame
* @param targetTimestamp The timestamp at which to interpolate
* @returns A new interpolated frame
*/
static interpolateFrames(
frame1: NonIndexedLeRobotDatasetRow,
frame2: NonIndexedLeRobotDatasetRow,
targetTimestamp: number
): NonIndexedLeRobotDatasetRow {
if (
targetTimestamp < frame1.timestamp ||
targetTimestamp > frame2.timestamp
) {
throw new Error(
"Target timestamp must be between the timestamps of the two frames"
);
}
const timeRange = frame2.timestamp - frame1.timestamp;
const interpolationFactor =
(targetTimestamp - frame1.timestamp) / timeRange;
// Interpolate action array
const interpolatedAction = LeRobotEpisode.interpolateArrays(
frame1.action,
frame2.action,
interpolationFactor
);
// Interpolate observation.state array
const interpolatedObservationState = LeRobotEpisode.interpolateArrays(
frame1["observation.state"],
frame2["observation.state"],
interpolationFactor
);
// Create the interpolated frame
return {
timestamp: targetTimestamp,
action: interpolatedAction,
"observation.state": interpolatedObservationState,
episode_index: frame1.episode_index,
task_index: frame1.task_index,
// Optional properties are not interpolated
frame_index: frame1.frame_index,
index: frame1.index,
};
}
/**
* Helper method to interpolate between two arrays
*
* @param array1 First array of values
* @param array2 Second array of values
* @param factor Interpolation factor (0-1)
* @returns Interpolated array
*/
private static interpolateArrays(
array1: any,
array2: any,
factor: number
): any {
// Handle different types of inputs
if (Array.isArray(array1) && Array.isArray(array2)) {
// For arrays, interpolate each element
return array1.map((value, index) => {
return value + (array2[index] - value) * factor;
});
} else if (typeof array1 === "object" && typeof array2 === "object") {
// For objects, interpolate each property
const result: any = {};
for (const key of Object.keys(array1)) {
if (key in array2) {
result[key] = array1[key] + (array2[key] - array1[key]) * factor;
} else {
result[key] = array1[key];
}
}
return result;
} else {
// For primitive values
return array1 + (array2 - array1) * factor;
}
}
}
/**
* Base interface for LeRobot dataset rows with common fields
*/
export interface NonIndexedLeRobotDatasetRow {
timestamp: number;
action: LeRobotAction;
"observation.state": LeRobotAction;
// properties are optional for back-converstion from normal rows
episode_index: number;
task_index: number;
frame_index?: number;
index?: number;
}
/**
* Represents a complete row in the LeRobot dataset format after indexing
* Used in the final exported dataset
*/
export interface LeRobotDatasetRow extends NonIndexedLeRobotDatasetRow {
frame_index: number;
index: number;
}
/**
* A mechanism to store and record, the video of all associated cameras
* as well as the teleoperator data
*
* follows the lerobot dataset format https://github.com/huggingface/lerobot/blob/cf86b9300dc83fdad408cfe4787b7b09b55f12cf/README.md#the-lerobotdataset-format
*/
export class LeRobotDatasetRecorder {
teleoperators: WebTeleoperator[];
videoStreams: { [key: string]: MediaStream };
mediaRecorders: { [key: string]: MediaRecorder };
videoChunks: { [key: string]: Blob[] };
videoBlobs: { [key: string]: Blob };
teleoperatorData: LeRobotEpisode[];
private _isRecording: boolean;
private episodeIndex: number = 0;
private taskIndex: number = 0;
fps: number;
taskDescription: string;
/**
* Ensures BlobPart compatibility across environments by converting Uint8Array
* to an ArrayBuffer with correct bounds and ArrayBuffer typing.
*/
private static toArrayBuffer(uint8: Uint8Array): ArrayBuffer {
const buffer = uint8.buffer;
if (buffer instanceof ArrayBuffer) {
return buffer.slice(
uint8.byteOffset,
uint8.byteOffset + uint8.byteLength
);
}
// Handle SharedArrayBuffer case by copying to ArrayBuffer
const arrayBuffer = new ArrayBuffer(uint8.byteLength);
new Uint8Array(arrayBuffer).set(uint8);
return arrayBuffer;
}
constructor(
teleoperators: WebTeleoperator[],
videoStreams: { [key: string]: MediaStream },
fps: number,
taskDescription: string = "Default task description"
) {
this.teleoperators = [];
if (teleoperators.length > 1)
throw Error(`
Currently, only 1 teleoperator can be recorded at a time!
Note : Do not attempt to create 2 different recorders via 2 different teleoperators, this would not work either
`);
this.addTeleoperator(teleoperators[0]);
this.mediaRecorders = {};
this.videoChunks = {};
this.videoBlobs = {};
this.videoStreams = {};
this.teleoperatorData = [];
this._isRecording = false;
this.fps = fps;
this.taskDescription = taskDescription;
for (const [key, stream] of Object.entries(videoStreams)) {
this.addVideoStream(key, stream);
}
}
get isRecording(): boolean {
return this._isRecording;
}
get currentEpisode(): LeRobotEpisode | undefined {
return this.teleoperatorData.at(-1);
}
/**
* Adds a new video stream to be recorded
* @param key The key to identify this video stream
* @param stream The media stream to record from
*/
addVideoStream(key: string, stream: MediaStream) {
console.log("Adding video stream", key);
if (this._isRecording) {
throw new Error("Cannot add video streams while recording");
}
// Add to video streams dictionary
this.videoStreams[key] = stream;
// Initialize MediaRecorder for this stream
this.mediaRecorders[key] = new MediaRecorder(stream, {
mimeType: "video/mp4",
});
// add a video chunk array for this stream
this.videoChunks[key] = [];
}
/**
* Add a new teleoperator and set up state update callbacks
* for recording joint position data in the LeRobot dataset format
*
* @param teleoperator The teleoperator to add callbacks to
*/
addTeleoperator(teleoperator: WebTeleoperator) {
teleoperator.addOnStateUpdateCallback((params) => {
if (this._isRecording) {
if (!this.currentEpisode)
throw Error(
"There is no current episode while recording, something is wrong!, this means that no frames exist on the recorder for some reason"
);
// Create a frame with the current state data
// Using the normalized configs for consistent data ranges
const frame: NonIndexedLeRobotDatasetRow = {
timestamp: params.commandSentTimestamp,
// For observation state, use the current motor positions
"observation.state": this.convertMotorConfigToArray(
params.newMotorConfigsNormalized
),
// For action, use the target positions that were commanded
action: this.convertMotorConfigToArray(
params.previousMotorConfigsNormalized
),
episode_index: this.episodeIndex,
task_index: this.taskIndex,
};
// Add the frame to the current episode
this.currentEpisode.add(frame);
}
});
this.teleoperators.push(teleoperator);
}
/**
* Starts recording for all teleoperators and video streams
*/
startRecording() {
console.log("Starting recording");
if (this._isRecording) {
console.warn("Recording already in progress");
return;
}
this._isRecording = true;
// add a new episode
this.teleoperatorData.push(new LeRobotEpisode());
// Start recording video streams
Object.entries(this.videoStreams).forEach(([key, stream]) => {
// Create a media recorder for this stream
const mediaRecorder = new MediaRecorder(stream, {
mimeType: "video/mp4",
});
// Handle data available events
mediaRecorder.ondataavailable = (event) => {
console.log("data available for", key);
if (event.data && event.data.size > 0) {
this.videoChunks[key].push(event.data);
}
};
// Save the recorder and start recording
this.mediaRecorders[key] = mediaRecorder;
mediaRecorder.start(1000); // Capture in 1-second chunks
console.log(`Started recording video stream: ${key}`);
});
}
setEpisodeIndex(index: number): void {
this.episodeIndex = index;
}
setTaskIndex(index: number): void {
this.taskIndex = index;
}
/**
* teleoperatorData by default only contains data
* for the episodes in a non-regularized manner
*
* this function returns episodes in a regularized manner, wherein
* the frames in each are interpolated through so that all frames are spaced
* equally through each other
*/
get episodes(): LeRobotEpisode[] {
const regularizedEpisodes: LeRobotEpisode[] = [];
let lastFrameIndex = 0;
for (let i = 0; i < this.teleoperatorData.length; i++) {
let episode = this.teleoperatorData[i];
const regularizedEpisode = episode.getInterpolatedRegularEpisode(
this.fps,
lastFrameIndex
);
regularizedEpisodes.push(regularizedEpisode);
lastFrameIndex += regularizedEpisode.frames?.at(-1)?.index || 0;
}
return regularizedEpisodes;
}
/**
* Stops recording for all teleoperators and video streams
* @returns An object containing teleoperator data and video blobs
*/
async stopRecording() {
if (!this._isRecording) {
console.warn("No recording in progress");
return { teleoperatorData: [], videoBlobs: {} };
}
this._isRecording = false;
// Stop all media recorders
const stopPromises = Object.entries(this.mediaRecorders).map(
([key, recorder]) => {
return new Promise<void>((resolve) => {
// Only do this if the recorder is active
if (recorder.state === "inactive") {
resolve();
return;
}
// When the recorder stops, create a blob
recorder.onstop = () => {
// Combine all chunks into a single blob
const chunks = this.videoChunks[key] || [];
const blob = new Blob(chunks, { type: "video/mp4" });
this.videoBlobs[key] = blob;
resolve();
};
// Stop the recorder
recorder.stop();
});
}
);
// Wait for all recorders to stop
await Promise.all(stopPromises);
return {
teleoperatorData: this.episodes,
videoBlobs: this.videoBlobs,
};
}
/**
* Clears the teleoperator data and video blobs
*/
clearRecording() {
this.teleoperatorData = [];
this.videoBlobs = {};
}
/**
* Action is a dictionary of motor positions, timestamp1 and timestamp2 are when the actions occurred
* reqTimestamp must be between timestamp1 and timestamp2
*
* the keys in action1 and action2 must match, this will loop through the dictionary
* and interpolate each action to the required timestamp
*
* @param action1 Motor positions at timestamp1
* @param action2 Motor positions at timestamp2
* @param timestamp1 The timestamp of action1
* @param timestamp2 The timestamp of action2
* @param reqTimestamp The timestamp at which to interpolate
* @returns The interpolated action
*/
_actionInterpolatate(
action1: any,
action2: any,
timestamp1: number,
timestamp2: number,
reqTimestamp: number
): any {
if (reqTimestamp < timestamp1 || reqTimestamp > timestamp2)
throw new Error("reqTimestamp must be between timestamp1 and timestamp2");
if (timestamp2 < timestamp1)
throw new Error("timestamp2 must be greater than timestamp1");
const numActions = Object.keys(action1).length;
const interpolatedAction: any = {};
const timeRange = timestamp2 - timestamp1;
for (let i = 0; i < numActions; i++) {
const key = Object.keys(action1)[i];
interpolatedAction[key] =
action1[key] +
((action2[key] - action1[key]) * (reqTimestamp - timestamp1)) /
timeRange;
}
return interpolatedAction;
}
/**
* Converts an action object to an array of numbers
* follows the same pattern as https://huggingface.co/datasets/lerobot/svla_so100_pickplace
* I am not really sure if the array can be in a different order
* but I am not going to risk it tbh 😛
*
* @param action The action object to convert
* @returns An array of numbers
*/
convertActionToArray(action: any): number[] {
return [
action["shoulder_pan"],
action["shoulder_lift"],
action["elbow_flex"],
action["wrist_flex"],
action["wrist_roll"],
action["gripper"],
];
}
/**
* Converts an array of MotorConfig objects to an action object
* following the same joint order as convertActionToArray
*
* @param motorConfigs Array of MotorConfig objects
* @returns An action object with joint positions
*/
convertMotorConfigToArray(motorConfigs: MotorConfig[]): number[] {
// Create a map for quick lookup of motor positions by name
const motorMap: Record<string, number> = {};
for (const config of motorConfigs) {
motorMap[config.name] = config.currentPosition;
}
// Define required joint names
const requiredJoints = [
"shoulder_pan",
"shoulder_lift",
"elbow_flex",
"wrist_flex",
"wrist_roll",
"gripper",
];
// Check that all required joints are present
const missingJoints = requiredJoints.filter(
(joint) => motorMap[joint] === undefined
);
if (missingJoints.length > 0) {
throw new Error(
`Missing required joints in motor configs: ${missingJoints.join(
", "
)}. Available joints: ${Object.keys(motorMap).join(", ")}`
);
}
// Return in the same order as convertActionToArray
return [
motorMap["shoulder_pan"],
motorMap["shoulder_lift"],
motorMap["elbow_flex"],
motorMap["wrist_flex"],
motorMap["wrist_roll"],
motorMap["gripper"],
];
}
/**
* Finds the closest timestamp to the target timestamp
*
* the data must have timestamps in ascending order
* uses binary search to get the closest timestamp
*
* @param data The data to search through
* @param targetTimestamp The target timestamp
* @returns The closest timestamp in the data's index
*/
_findClosestTimestampBefore(data: any[], targetTimestamp: number): number {
let firstIndex = 0;
let lastIndex = data.length - 1;
while (firstIndex <= lastIndex) {
const middleIndex = Math.floor((firstIndex + lastIndex) / 2);
const middleTimestamp = data[middleIndex].timestamp;
if (middleTimestamp === targetTimestamp) {
return middleIndex;
} else if (middleTimestamp < targetTimestamp) {
firstIndex = middleIndex + 1;
} else {
lastIndex = middleIndex - 1;
}
}
return lastIndex;
}
/**
* Takes non-regularly spaced lerobot-ish data and interpolates it to a regularly spaced dataset
* also adds additional
* - frame_index
* - episode_index
* - index columns
*
* to match lerobot dataset requirements
*/
_interpolateAndCompleteLerobotData(
fps: number,
frameData: NonIndexedLeRobotDatasetRow[],
lastFrameIndex: number
): LeRobotDatasetRow[] {
const interpolatedData: LeRobotDatasetRow[] = [];
const minTimestamp = frameData[0].timestamp;
const maxTimestamp = frameData[frameData.length - 1].timestamp;
const timeDifference = maxTimestamp - minTimestamp;
const numFrames = Math.floor(timeDifference * fps);
const firstFrame = frameData[0];
console.log(
"frames before interpolation",
numFrames,
frameData[0].timestamp,
frameData[frameData.length - 1].timestamp,
fps
);
interpolatedData.push({
timestamp: firstFrame.timestamp,
action: this.convertActionToArray(firstFrame.action),
"observation.state": this.convertActionToArray(
firstFrame["observation.state"]
),
episode_index: firstFrame.episode_index,
task_index: firstFrame.task_index,
frame_index: 0,
index: lastFrameIndex,
});
// start from 1 as the first frame is pushed already (see above)
for (let i = 1; i < numFrames; i++) {
const timestamp = i / fps;
const closestIndex = this._findClosestTimestampBefore(
frameData,
timestamp
);
const nextIndex = closestIndex + 1;
const closestItemData = frameData[closestIndex];
const nextItemData = frameData[nextIndex];
const action = this._actionInterpolatate(
closestItemData.action,
nextItemData.action,
closestItemData.timestamp,
nextItemData.timestamp,
timestamp
);
const observation_state = this._actionInterpolatate(
closestItemData["observation.state"],
nextItemData["observation.state"],
closestItemData.timestamp,
nextItemData.timestamp,
timestamp
);
interpolatedData.push({
timestamp: timestamp,
action: this.convertActionToArray(action),
"observation.state": this.convertActionToArray(observation_state),
episode_index: closestItemData.episode_index,
task_index: closestItemData.task_index,
frame_index: i,
index: lastFrameIndex + i,
});
}
return interpolatedData;
}
/**
* converts all the frames of a recording into lerobot dataset frame style
*
* NOTE : This does not interpolate the data, you are only working with raw data
* that is called by the teleop when things are actively "changing"
* @param episodeRough
*/
_convertToLeRobotDataFormatFrames(
episodeRough: any[]
): NonIndexedLeRobotDatasetRow[] {
const properFormatFrames: NonIndexedLeRobotDatasetRow[] = [];
const firstTimestamp = episodeRough[0].commandSentTimestamp;
for (let i = 0; i < episodeRough.length; i++) {
const frameRough = episodeRough[i];
properFormatFrames.push({
timestamp: frameRough.commandSentTimestamp - firstTimestamp, // so timestamps start from 0, and are in seconds
action: frameRough.previousMotorConfigsNormalized,
"observation.state": frameRough.newMotorConfigsNormalized,
episode_index: frameRough.episodeIndex,
task_index: frameRough.taskIndex,
});
}
return properFormatFrames;
}
/**
* Converts teleoperator data to a parquet blob
* @private
* @returns Array of objects containing parquet file content and path
*/
private async _exportEpisodesToBlob(
episodes: LeRobotEpisode[]
): Promise<{ content: Blob; path: string }[]> {
// combine all the frames
let data: NonIndexedLeRobotDatasetRow[] = [];
const episodeBlobs: any[] = [];
for (let i = 0; i < episodes.length; i++) {
const episode = episodes[i];
data = episode.frames;
const { tableFromArrays, vectorFromArray } = arrow;
const timestamps = data.map((row: any) => row.timestamp);
const actions = data.map((row: any) => row.action);
const observationStates = data.map(
(row: any) => row["observation.state"]
);
const episodeIndexes = data.map((row: any) => row.episode_index);
const taskIndexes = data.map((row: any) => row.task_index);
const frameIndexes = data.map((row: any) => row.frame_index);
const indexes = data.map((row: any) => row.index);
const table = tableFromArrays({
timestamp: timestamps,
// @ts-ignore, this works, idk why
action: vectorFromArray(
actions,
new arrow.List(new arrow.Field("item", new arrow.Float32()))
),
// @ts-ignore, this works, idk why
"observation.state": vectorFromArray(
observationStates,
new arrow.List(new arrow.Field("item", new arrow.Float32()))
),
episode_index: episodeIndexes,
task_index: taskIndexes,
frame_index: frameIndexes,
index: indexes,
});
const wasmUrl =
"https://cdn.jsdelivr.net/npm/parquet-wasm@0.6.1/esm/parquet_wasm_bg.wasm";
const initWasm = parquet.default;
await initWasm(wasmUrl);
const wasmTable = parquet.Table.fromIPCStream(
arrow.tableToIPC(table, "stream")
);
const writerProperties = new parquet.WriterPropertiesBuilder()
.setCompression(parquet.Compression.UNCOMPRESSED)
.build();
const parquetUint8Array = parquet.writeParquet(
wasmTable,
writerProperties
);
const numpadded = i.toString().padStart(6, "0");
const content = new Blob([
LeRobotDatasetRecorder.toArrayBuffer(parquetUint8Array as Uint8Array),
]);
episodeBlobs.push({
content,
path: `data/chunk-000/episode_${numpadded}.parquet`,
});
}
return episodeBlobs;
}
/**
* Exports the teleoperator data in lerobot format
* @param format The format to return the data in ('json' or 'blob')
* @returns Either an array of data objects or a Uint8Array blob depending on format
*/
exportEpisodes(format: "json" | "blob" = "json") {
if (this._isRecording)
throw new Error("This can only be called after recording has stopped!");
const data = this.episodes;
if (format === "json") {
return data;
} else {
return this._exportEpisodesToBlob(data);
}
}
/**
* Exports the media (video) data as blobs
* @returns A dictionary of video blobs with the same keys as videoStreams
*/
async exportMediaData(): Promise<{ [key: string]: Blob }> {
if (this._isRecording)
throw new Error("This can only be called after recording has stopped!");
return this.videoBlobs;
}
/**
* Generates metadata for the dataset
* @returns Metadata object for the LeRobot dataset
*/
async generateMetadata(data: any[]): Promise<any> {
// Calculate total episodes, frames, and tasks
let total_episodes = 0;
const total_frames = data.length;
let total_tasks = 0;
for (const row of data) {
total_episodes = Math.max(total_episodes, row.episode_index);
total_tasks = Math.max(total_tasks, row.task_index);
}
// Create video info objects for each video stream
const videos_info: VideoInfo[] = Object.keys(this.videoBlobs).map((key) => {
// Default values - in a production environment, you would extract
// these from the actual video metadata using the key to identify the video source
console.log(`Generating metadata for video stream: ${key}`);
return {
height: 480,
width: 640,
channels: 3,
codec: "h264",
pix_fmt: "yuv420p",
is_depth_map: false,
has_audio: false,
};
});
// Calculate approximate file sizes in MB
const data_files_size_in_mb = Math.round(data.length * 0.001); // Estimate
// Calculate video size by summing the sizes of all video blobs and converting to MB
let video_files_size_in_mb = 0;
for (const blob of Object.values(this.videoBlobs)) {
video_files_size_in_mb += blob.size / (1024 * 1024);
}
video_files_size_in_mb = Math.round(video_files_size_in_mb);
// Generate and return the metadata
return getMetadataInfo({
total_episodes,
total_frames,
total_tasks,
chunks_size: 1000, // Default chunk size
fps: this.fps,
splits: { train: `0:${total_episodes}` }, // All episodes in train split
features: {}, // Additional features can be added here
videos_info,
data_files_size_in_mb,
video_files_size_in_mb,
});
}
/**
* Generates statistics for the dataset
* @returns Statistics object for the LeRobot dataset
*/
async getStatistics(data: any[]): Promise<any> {
// Get camera keys from the video blobs
const cameraKeys = Object.keys(this.videoBlobs);
// Generate stats using the data and camera keys
return getStats(data, cameraKeys);
}
/**
* Creates a tasks.parquet file containing task description
* @returns A Uint8Array blob containing the parquet data
*/
async createTasksParquet(): Promise<Uint8Array> {
// Create a simple data structure with the task description
const tasksData = [
{
task_index: 0,
__index_level_0__: this.taskDescription,
},
];
// Create Arrow table from the data
const taskIndexArr = arrow.vectorFromArray(
tasksData.map((d) => d.task_index),
new arrow.Int32()
);
const descriptionArr = arrow.vectorFromArray(
tasksData.map((d) => d.__index_level_0__),
new arrow.Utf8()
);
const table = arrow.tableFromArrays({
// @ts-ignore, this works, idk why
task_index: taskIndexArr,
// @ts-ignore, this works, idk why
__index_level_0__: descriptionArr,
});
// Initialize the WASM module
const wasmUrl =
"https://cdn.jsdelivr.net/npm/parquet-wasm@0.6.1/esm/parquet_wasm_bg.wasm";
const initWasm = parquet.default;
await initWasm(wasmUrl);
// Convert Arrow table to Parquet WASM table
const wasmTable = parquet.Table.fromIPCStream(
arrow.tableToIPC(table, "stream")
);
// Set compression properties
const writerProperties = new parquet.WriterPropertiesBuilder()
.setCompression(parquet.Compression.UNCOMPRESSED)
.build();
// Write the Parquet file
return parquet.writeParquet(wasmTable, writerProperties);
}
/**
* Creates the episodes statistics parquet file
* @returns A Uint8Array blob containing the parquet data
*/
async getEpisodeStatistics(data: any[]): Promise<Uint8Array> {
const { vectorFromArray } = arrow;
const statistics = await this.getStatistics(data);
// Calculate total episodes and frames
let total_episodes = 0;
for (let row of data) {
total_episodes = Math.max(total_episodes, row.episode_index);
}
total_episodes += 1; // +1 since episodes start from 0
const episodes: any[] = [];
// we'll create one row per episode
for (
let episode_index = 0;
episode_index < total_episodes;
episode_index++
) {
// Get data for this episode only
const episodeData = data.filter(
(row) => row.episode_index === episode_index
);
// Extract timestamps for this episode
const timestamps = episodeData.map((row) => row.timestamp);
let min_timestamp = Infinity;
let max_timestamp = -Infinity;
for (let timestamp of timestamps) {
min_timestamp = Math.min(min_timestamp, timestamp);
max_timestamp = Math.max(max_timestamp, timestamp);
}
// Camera keys from video blobs
const cameraKeys = Object.keys(this.videoBlobs);
// Create entry for this episode
const episodeEntry: any = {
// Basic episode information
episode_index: episode_index,
"data/chunk_index": 0,
"data/file_index": 0,
dataset_from_index: 0,
dataset_to_index: episodeData.length - 1,
length: episodeData.length,
tasks: [0], // Task index 0, could be extended for multiple tasks
// Meta information
"meta/episodes/chunk_index": 0,
"meta/episodes/file_index": 0,
};
// Add video information for each camera
cameraKeys.forEach((key) => {
episodeEntry[`videos/observation.images.${key}/chunk_index`] = 0;
episodeEntry[`videos/observation.images.${key}/file_index`] = 0;
episodeEntry[`videos/observation.images.${key}/from_timestamp`] =
min_timestamp;
episodeEntry[`videos/observation.images.${key}/to_timestamp`] =
max_timestamp;
});
// Add statistics for each field
// This is a simplified approach - in a real implementation, you'd calculate
// these values for each episode individually
// Add timestamp statistics
episodeEntry["stats/timestamp/min"] = [statistics.timestamp.min];
episodeEntry["stats/timestamp/max"] = [statistics.timestamp.max];
episodeEntry["stats/timestamp/mean"] = [statistics.timestamp.mean];
episodeEntry["stats/timestamp/std"] = [statistics.timestamp.std];
episodeEntry["stats/timestamp/count"] = [statistics.timestamp.count];
// Add frame_index statistics
episodeEntry["stats/frame_index/min"] = [statistics.frame_index.min];
episodeEntry["stats/frame_index/max"] = [statistics.frame_index.max];
episodeEntry["stats/frame_index/mean"] = [statistics.frame_index.mean];
episodeEntry["stats/frame_index/std"] = [statistics.frame_index.std];
episodeEntry["stats/frame_index/count"] = [statistics.frame_index.count];
// Add episode_index statistics
episodeEntry["stats/episode_index/min"] = [statistics.episode_index.min];
episodeEntry["stats/episode_index/max"] = [statistics.episode_index.max];
episodeEntry["stats/episode_index/mean"] = [
statistics.episode_index.mean,
];
episodeEntry["stats/episode_index/std"] = [statistics.episode_index.std];
episodeEntry["stats/episode_index/count"] = [
statistics.episode_index.count,
];
// Add task_index statistics
episodeEntry["stats/task_index/min"] = [statistics.task_index.min];
episodeEntry["stats/task_index/max"] = [statistics.task_index.max];
episodeEntry["stats/task_index/mean"] = [statistics.task_index.mean];
episodeEntry["stats/task_index/std"] = [statistics.task_index.std];
episodeEntry["stats/task_index/count"] = [statistics.task_index.count];
// Add index statistics
episodeEntry["stats/index/min"] = [0];
episodeEntry["stats/index/max"] = [episodeData.length - 1];
episodeEntry["stats/index/mean"] = [episodeData.length / 2];
episodeEntry["stats/index/std"] = [episodeData.length / 4]; // Approximate std
episodeEntry["stats/index/count"] = [episodeData.length];
// Add action statistics (placeholder)
episodeEntry["stats/action/min"] = [0.0];
episodeEntry["stats/action/max"] = [1.0];
episodeEntry["stats/action/mean"] = [0.5];
episodeEntry["stats/action/std"] = [0.2];
episodeEntry["stats/action/count"] = [episodeData.length];
// Add observation.state statistics (placeholder)
episodeEntry["stats/observation.state/min"] = [0.0];
episodeEntry["stats/observation.state/max"] = [1.0];
episodeEntry["stats/observation.state/mean"] = [0.5];
episodeEntry["stats/observation.state/std"] = [0.2];
episodeEntry["stats/observation.state/count"] = [episodeData.length];
// Add observation.images statistics for each camera
cameraKeys.forEach((key) => {
// Get the image statistics from the overall statistics
const imageStats = statistics[`observation.images.${key}`] || {
min: [[[0.0]], [[0.0]], [[0.0]]],
max: [[[255.0]], [[255.0]], [[255.0]]],
mean: [[[127.5]], [[127.5]], [[127.5]]],
std: [[[50.0]], [[50.0]], [[50.0]]],
count: [[[episodeData.length * 3]]],
};
episodeEntry[`stats/observation.images.${key}/min`] = imageStats.min;
episodeEntry[`stats/observation.images.${key}/max`] = imageStats.max;
episodeEntry[`stats/observation.images.${key}/mean`] = imageStats.mean;
episodeEntry[`stats/observation.images.${key}/std`] = imageStats.std;
episodeEntry[`stats/observation.images.${key}/count`] =
imageStats.count;
});
episodes.push(episodeEntry);
}
// Create vector arrays for each column
const columns: any = {};
// Define column names and default types
const columnNames = [
"episode_index",
"data/chunk_index",
"data/file_index",
"dataset_from_index",
"dataset_to_index",
"length",
"meta/episodes/chunk_index",
"meta/episodes/file_index",
"tasks",
];
// Add camera-specific columns
const cameraKeys = Object.keys(this.videoBlobs);
cameraKeys.forEach((key) => {
columnNames.push(
`videos/observation.images.${key}/chunk_index`,
`videos/observation.images.${key}/file_index`,
`videos/observation.images.${key}/from_timestamp`,
`videos/observation.images.${key}/to_timestamp`
);
});
// Add statistic columns for each field
const statFields = [
"timestamp",
"frame_index",
"episode_index",
"task_index",
"index",
"action",
"observation.state",
];
statFields.forEach((field) => {
columnNames.push(
`stats/${field}/min`,
`stats/${field}/max`,
`stats/${field}/mean`,
`stats/${field}/std`,
`stats/${field}/count`
);
});
// Add image statistic columns for each camera
cameraKeys.forEach((key) => {
columnNames.push(
`stats/observation.images.${key}/min`,
`stats/observation.images.${key}/max`,
`stats/observation.images.${key}/mean`,
`stats/observation.images.${key}/std`,
`stats/observation.images.${key}/count`
);
});
// Create vector arrays for each column
columnNames.forEach((columnName) => {
const values = episodes.map((ep) => ep[columnName] || 0);
// Check if the column is an array type and needs special handling
if (columnName.includes("stats/") || columnName === "tasks") {
// Handle different types of array columns based on their naming pattern
if (columnName.includes("/count")) {
// Bigint arrays for count fields
// @ts-ignore
columns[columnName] = vectorFromArray(
values.map((v) => Number(v)),
new arrow.List(new arrow.Field("item", new arrow.Int64()))
);
} else if (
columnName.includes("/min") ||
columnName.includes("/max") ||
columnName.includes("/mean") ||
columnName.includes("/std")
) {
// Double arrays for min, max, mean, std fields
if (
columnName.includes("observation.images") &&
(columnName.includes("/min") ||
columnName.includes("/max") ||
columnName.includes("/mean") ||
columnName.includes("/std"))
) {
// These are 3D arrays [[[value]]]
// For 3D arrays, we need nested Lists
// @ts-ignore
columns[columnName] = vectorFromArray(
values,
new arrow.List(
new arrow.Field(
"item",
new arrow.List(
new arrow.Field(
"subitem",
new arrow.List(
new arrow.Field("value", new arrow.Float64())
)
)
)
)
)
);
} else {
// These are normal arrays [value]
// @ts-ignore
columns[columnName] = vectorFromArray(
values,
new arrow.List(new arrow.Field("item", new arrow.Float64()))
);
}
} else {
// Default to Float64 List for other array types
// @ts-ignore
columns[columnName] = vectorFromArray(
values,
new arrow.List(new arrow.Field("item", new arrow.Float64()))
);
}
} else {
// For non-array columns, use regular vectorFromArray
// @ts-ignore
columns[columnName] = vectorFromArray(values);
}
});
// Create the table with all columns
const table = arrow.tableFromArrays(columns);
// Initialize the WASM module
const wasmUrl =
"https://cdn.jsdelivr.net/npm/parquet-wasm@0.6.1/esm/parquet_wasm_bg.wasm";
const initWasm = parquet.default;
await initWasm(wasmUrl);
// Convert Arrow table to Parquet WASM table
const wasmTable = parquet.Table.fromIPCStream(
arrow.tableToIPC(table, "stream")
);
// Set compression properties
const writerProperties = new parquet.WriterPropertiesBuilder()
.setCompression(parquet.Compression.UNCOMPRESSED)
.build();
// Write the Parquet file
return parquet.writeParquet(wasmTable, writerProperties);
}
generateREADME(metaInfo: string) {
return generateREADME(metaInfo);
}
/**
* Creates an array of path and blob content objects for the LeRobot dataset
*
* @returns An array of {path, content} objects representing the dataset files
* @private
*/
async _exportForLeRobotBlobs() {
const teleoperatorDataJson = (await this.exportEpisodes("json")) as any[];
const parquetEpisodeDataFiles = await this._exportEpisodesToBlob(
teleoperatorDataJson
);
const videoBlobs = await this.exportMediaData();
const metadata = await this.generateMetadata(teleoperatorDataJson);
const statistics = await this.getStatistics(teleoperatorDataJson);
const tasksParquet = await this.createTasksParquet();
const episodesParquet = await this.getEpisodeStatistics(
teleoperatorDataJson
);
const readme = this.generateREADME(JSON.stringify(metadata));
// Create the blob array with proper paths
const blobArray = [
...parquetEpisodeDataFiles,
{
path: "meta/info.json",
content: new Blob([JSON.stringify(metadata, null, 2)], {
type: "application/json",
}),
},
{
path: "meta/stats.json",
content: new Blob([JSON.stringify(statistics, null, 2)], {
type: "application/json",
}),
},
{
path: "meta/tasks.parquet",
content: new Blob([
LeRobotDatasetRecorder.toArrayBuffer(tasksParquet as Uint8Array),
]),
},
{
path: "meta/episodes/chunk-000/file-000.parquet",
content: new Blob([
LeRobotDatasetRecorder.toArrayBuffer(episodesParquet as Uint8Array),
]),
},
{
path: "README.md",
content: new Blob([readme], { type: "text/markdown" }),
},
];
// Add video blobs with proper paths
for (const [key, blob] of Object.entries(videoBlobs)) {
blobArray.push({
path: `videos/chunk-000/observation.images.${key}/episode_000000.mp4`,
content: blob,
});
}
return blobArray;
}
/**
* Creates a ZIP file from the dataset blobs
*
* @returns A Blob containing the ZIP file
* @private
*/
async _exportForLeRobotZip() {
const blobArray = await this._exportForLeRobotBlobs();
const zip = new JSZip();
// Add all blobs to the zip with their paths
for (const item of blobArray) {
// Split the path to handle directories
const pathParts = item.path.split("/");
const fileName = pathParts.pop() || "";
let currentFolder = zip;
// Create nested folders as needed
if (pathParts.length > 0) {
for (const part of pathParts) {
currentFolder = currentFolder.folder(part) || currentFolder;
}
}
// Add file to the current folder
currentFolder.file(fileName, item.content);
}
// Generate the zip file
return await zip.generateAsync({ type: "blob" });
}
/**
* Uploads the LeRobot dataset to Hugging Face
*
* @param username Hugging Face username
* @param repoName Repository name for the dataset
* @param accessToken Hugging Face access token
* @returns The LeRobotHFUploader instance used for upload
*/
async _exportForLeRobotHuggingface(
username: string,
repoName: string,
accessToken: string
) {
// Create the blobs array for upload
const blobArray = await this._exportForLeRobotBlobs();
// Create the uploader
const uploader = new LeRobotHFUploader(username, repoName);
// Convert blobs to File objects for HF uploader
const files = blobArray.map((item) => {
return {
path: item.path,
content: item.content,
};
});
// Generate a unique reference ID for tracking the upload
const referenceId = `lerobot-upload-${Date.now()}`;
try {
// Start the upload process
uploader.createRepoAndUploadFiles(files, accessToken, referenceId);
console.log(`Successfully uploaded dataset to ${username}/${repoName}`);
return uploader;
} catch (error) {
console.error("Error uploading to Hugging Face:", error);
throw error;
}
}
/**
* Uploads the LeRobot dataset to Amazon S3
*
* @param bucketName S3 bucket name
* @param accessKeyId AWS access key ID
* @param secretAccessKey AWS secret access key
* @param region AWS region (default: us-east-1)
* @param prefix Optional prefix (folder) to upload files to within the bucket
* @returns The LeRobotS3Uploader instance used for upload
*/
async _exportForLeRobotS3(
bucketName: string,
accessKeyId: string,
secretAccessKey: string,
region: string = "us-east-1",
prefix: string = ""
) {
// Create the blobs array for upload
const blobArray = await this._exportForLeRobotBlobs();
// Create the uploader
const uploader = new LeRobotS3Uploader(bucketName, region);
// Convert blobs to File objects for S3 uploader
const files = blobArray.map((item) => {
return {
path: item.path,
content: item.content,
};
});
// Generate a unique reference ID for tracking the upload
const referenceId = `lerobot-s3-upload-${Date.now()}`;
try {
// Start the upload process
uploader.checkBucketAndUploadFiles(
files,
accessKeyId,
secretAccessKey,
prefix,
referenceId
);
console.log(`Successfully uploaded dataset to S3 bucket: ${bucketName}`);
return uploader;
} catch (error) {
console.error("Error uploading to S3:", error);
throw error;
}
}
/**
* Exports the LeRobot dataset in various formats
*
* @param format The export format - 'blobs', 'zip', 'zip-download', 'huggingface', or 's3'
* @param options Additional options for specific formats
* @param options.username Hugging Face username (if not provided for "huggingface" format, it will use the default username)
* @param options.repoName Hugging Face repository name (required for 'huggingface' format)
* @param options.accessToken Hugging Face access token (required for 'huggingface' format)
* @param options.bucketName S3 bucket name (required for 's3' format)
* @param options.accessKeyId AWS access key ID (required for 's3' format)
* @param options.secretAccessKey AWS secret access key (required for 's3' format)
* @param options.region AWS region (optional for 's3' format, default: us-east-1)
* @param options.prefix S3 prefix/folder (optional for 's3' format)
* @returns The exported data in the requested format or the uploader instance for 'huggingface'/'s3' formats
*/
async exportForLeRobot(
format:
| "blobs"
| "zip"
| "zip-download"
| "huggingface"
| "s3" = "zip-download",
options?: {
username?: string;
repoName?: string;
accessToken?: string;
bucketName?: string;
accessKeyId?: string;
secretAccessKey?: string;
region?: string;
prefix?: string;
}
) {
switch (format) {
case "blobs":
return this._exportForLeRobotBlobs();
case "zip":
return this._exportForLeRobotZip();
case "huggingface":
// Validate required options for Hugging Face upload
if (!options || !options.repoName || !options.accessToken) {
throw new Error(
"Hugging Face upload requires repoName, and accessToken options"
);
}
if (!options.username) {
const hub = await import("@huggingface/hub");
const { name: username } = await hub.whoAmI({
accessToken: options.accessToken,
});
options.username = username;
}
return this._exportForLeRobotHuggingface(
options.username,
options.repoName,
options.accessToken
);
case "s3":
// Validate required options for S3 upload
if (
!options ||
!options.bucketName ||
!options.accessKeyId ||
!options.secretAccessKey
) {
throw new Error(
"S3 upload requires bucketName, accessKeyId, and secretAccessKey options"
);
}
return this._exportForLeRobotS3(
options.bucketName,
options.accessKeyId,
options.secretAccessKey,
options.region,
options.prefix
);
case "zip-download":
default:
// Get the zip blob
const zipContent = await this._exportForLeRobotZip();
// Create a URL for the zip file
const url = URL.createObjectURL(zipContent);
// Create a download link and trigger the download
const link = document.createElement("a");
link.href = url;
link.download = `lerobot_dataset_${new Date()
.toISOString()
.replace(/[:.]/g, "-")}.zip`;
document.body.appendChild(link);
link.click();
// Clean up
setTimeout(() => {
document.body.removeChild(link);
URL.revokeObjectURL(url);
}, 100);
return zipContent;
}
}
}