shiyixia/robot-kinematics-dataset

🤖 EgoCentric Multi-Episode Kinematics Trajectory Dataset

This repository contains multi-episode egocentric (first-person) video recordings paired with extracted 3D hand, wrist, and arm kinematics trajectories. The dataset is specifically structured for Imitation Learning (IL), Behavior Cloning (BC), Isaac Sim Kinematics Replay, and Reinforcement Learning Datasets (RLDS) training pipelines.

---

📌 Dataset Overview

Source Domain	Egocentric / First-Person View Video Stream (GoPro / Ego-cam)
Extraction Engine	Dual-pipeline tracking via YOLO11m-Pose (Arm Tracking) and MediaPipe Hands (3D Hand Landmarks & Gripper Distance Estimation)
Target Robot Compatibility	Franka Emika Panda, Universal Robots, and Custom Parallel-Jaw End-Effectors
Data Formats	Apache Parquet, RLDS JSON, and Binary TFRecord (exported concurrently)

---

📁 Repository Structure

my-robot-kinematics-dataset/
├── README.md                       # Dataset documentation & Hugging Face Dataset Card
├── videos/
│   ├── episode_001_raw.mp4         # Raw egocentric input video (Episode 1)
│   ├── episode_001_kinematics.mp4  # Visual tracking overlay & skeleton visualization (Episode 1)
│   ├── episode_002_raw.mp4         # Raw egocentric input video (Episode 2)
│   └── episode_002_kinematics.mp4  # Visual tracking overlay & skeleton visualization (Episode 2)
└── data/
    ├── episode_001.parquet         # Columnar Parquet dataset for Episode 1 (PyTorch / LeRobot)
    ├── episode_001.json            # Human-readable RLDS JSON schema for Episode 1
    ├── episode_001.tfrecord        # Binary TFRecord format for Episode 1 (Open X-Embodiment)

---

📊 Data Schema Specifications

Each step in the dataset trajectories represents a sequential video frame entry with the following feature mappings:

Feature Field Name	Data Type	Range / Dimension	Description
step_index	int64	[1, N]	Frame sequence index within the episode
timestamp_sec	float64	≥ 0.0	Exact temporal frame timestamp in seconds
obs_shoulder_xy	list[float]	[X, Y] ∈ [0.0, 1.0]	Normalized 2D Shoulder joint coordinate
obs_elbow_xy	list[float]	[X, Y] ∈ [0.0, 1.0]	Normalized 2D Elbow joint coordinate
obs_wrist_xy	list[float]	[X, Y] ∈ [0.0, 1.0]	Normalized 2D Wrist joint coordinate
obs_gripper_width	float64	[0.0, 1.0]	Normalized gripper distance (0.0 = Closed, 1.0 = Open)
action_delta_wrist_xy	list[float]	[ΔX, ΔY]	Relative wrist displacement relative to previous frame
action_delta_gripper_width	float64	ΔG	Relative gripper state change relative to previous frame
task_grasp_phase	string	Categorical	Discrete state: OPEN, APPROACHING, or GRASPING
is_first	boolean	True / False	Episode boundary flag indicating starting frame
is_last	boolean	True / False	Episode boundary flag indicating terminating frame

---

💻 Quickstart Integration Examples

1. High-Speed Cloud Streaming via Pandas (Parquet)

You can load the dataset directly into a Pandas DataFrame using the hf:// protocol, without cloning the full repository:

import pandas as pd

# Stream Episode 001 directly from Hugging Face
parquet_url = "hf://datasets/shiyixia/robot-kinematics-dataset/data/episode_001.parquet"
df = pd.read_parquet(parquet_url)

print("--- Episode 001 Trajectory Preview ---")
print(df[["step_index", "obs_wrist_xy", "obs_gripper_width", "task_grasp_phase"]].head())

2. Multi-Episode Batch Aggregation for PyTorch DataLoaders

import pandas as pd

episodes = ["episode_001", "episode_002"]
dataset_frames = []

base_url = "hf://datasets/shiyixia/robot-kinematics-dataset/data"

for ep in episodes:
    url = f"{base_url}/{ep}.parquet"
    ep_df = pd.read_parquet(url)
    ep_df["episode_id"] = ep
    dataset_frames.append(ep_df)

full_dataset = pd.concat(dataset_frames, ignore_index=True)
print(f"✅ Total Processed Steps Across All Episodes: {len(full_dataset)}")

3. Loading RLDS Structured JSON in Python

import json
import requests

json_url = "https://huggingface.co/datasets/shiyixia/robot-kinematics-dataset/raw/main/data/episode_001.json"
response = requests.get(json_url)
episode_data = response.json()

print(f"Episode ID: {episode_data['episode_id']}")
print(f"Total Steps: {episode_data['total_steps']}")
print(f"First Step Action Delta: {episode_data['steps'][0]['action']['delta_wrist_xy']}")

---

📬 Contact & Custom Data Requests

If you are interested in purchasing any kind of skill-type video/kinematics datasets, or have suggestions for improving this dataset, please reach out:

📧 shiyixia2000@gmail.com

---

📜 Citation & License

If you utilize this dataset in your research, simulation experiments (NVIDIA Isaac Sim, MuJoCo, Gazebo), or robot learning benchmarks, please cite this repository:

@dataset{egograsp_kinematics_2026,
  author       = {Your Name/Organization},
  title        = {EgoCentric Multi-Episode Kinematics Trajectory Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/shiyixia/robot-kinematics-dataset}}
}

License: Distributed under the MIT License.