S0101_SIM_Leisaac_Orange

Overview

This repository contains the initial raw teleoperation dataset collected as part of a Physical AI end-to-end lifecycle project. The dataset captures human-in-the-loop teleoperation of a simulated robot arm performing an object manipulation task (orange picking and placing) inside a high-fidelity simulation environment.

The work builds on Lightwheel AI's Leisaac framework, running on top of NVIDIA Isaac Sim and Isaac Lab, and represents the data generation and validation phase of a broader pipeline that will later include domain randomization, large-scale synthetic data generation, and policy release.

This release focuses only on raw teleoperation data. Trained policy weights and fully domain-randomized datasets will be released at a later stage.

---

Project Goals

• Build an end-to-end Physical AI workflow
• Validate teleoperation-driven data collection in simulation
• Create high-quality, sim-ready datasets for robotic manipulation
• Explore realistic scene generation and domain randomization
• Train and evaluate policies fully in simulation before real-world transfer

---

Simulation Stack

The project uses the following stack:

• NVIDIA Isaac Sim: Physics-based robotics simulation
• NVIDIA Isaac Lab: Task and reinforcement learning framework
• Lightwheel AI Leisaac: High-level robotics workflows and teleoperation
• Marble World Labs (World Labs Marble AI): AI-generated, photorealistic simulation environments
• NVIDIA MimicGen: Post-teleoperation domain randomization and data scaling

Leisaac setup was tested both locally and via Docker deployment before data collection.

AI-Generated Environment Generation (Marble World Labs)

The simulation environment used in this project was generated using World Labs Marble AI, enabling rapid creation of high-fidelity, sim-ready 3D worlds suitable for robotic training.

Instead of manual scene construction, a single environment can be generated in minutes and reused to produce large-scale environmental variation. Properties such as lighting, layout, clutter, and textures can be randomized, making this approach highly effective for domain generalization in simulation-based robot learning.

This workflow and its motivation are documented in the following LinkedIn post:

👉 From Days of Scene Building to Minutes of World Generation for Robot Training

The approach was originally inspired by an NVIDIA Developer Blog on Isaac Sim, referenced in the post above.

---

Task Description

• Robot: S0101 robotic arm

• Task: Orange Picker

• Control Mode: Keyboard-based teleoperation

• Objective:

  • Navigate the arm to grasp an orange
  • Lift and transport the object
  • Place the orange onto a plate
  • Release the gripper cleanly

All demonstrations were recorded directly inside the simulation environment.

---

Dataset Contents

This dataset represents raw teleoperation rollouts and associated sensor streams recorded during task execution.

Data Includes

• Robot joint states
• End-effector poses
• Gripper open/close actions
• Object states (orange, plate)
• Simulation timestamps
• Action and observation trajectories

Data Annotation

After recording, clips and sensor streams were manually tagged to label key events such as:

• Grasping the orange
• Lifting and transport
• Placement over the plate
• Gripper opening and object release

These annotations are intended for downstream imitation learning, policy fine-tuning, and dataset filtering.

---

Repository Structure

S0101_SIM_Leisaac_Orange/
├── Marble World/
│   └── Kitchen/
│       ├── Rustic Kitchen with Natural Light.usdz
│       └── Rustic Kitchen with Natural Light_collider.glb
│
├── datasets/
│   └── dataset.hdf5
│
└── README.md

Structure Breakdown

• Marble World/

  • Contains AI-generated, sim-ready environment assets
  • High-fidelity kitchen scene used for teleoperation
  • Includes both visual and collider geometry

• datasets/dataset.hdf5

  • Main teleoperation dataset file
  • Stores raw trajectories, sensor data, and annotations

---

Physical AI Lifecycle Coverage

This repository corresponds to the early and middle stages of the Physical AI pipeline:

1. Simulation environment setup
2. Teleoperation-based data collection
3. Manual tagging and dataset validation
4. Initial policy training and evaluation in simulation

Later stages (not included in this release):

• Large-scale domain randomization
• Synthetic data expansion
• Final policy training and release
• Sim-to-real transfer experiments

---

Policy Training Summary (Not Released)

Beyond this dataset, the following steps were completed privately:

• Fine-tuning GROOT N1.5 using QLoRA
• Training based on documented Isaac and Leisaac workflows
• Policy inference and validation inside simulation

Model weights and training configs will be released separately.

---

Domain Randomization (Not Released)

Domain randomization was also performed privately as part of the broader Physical AI workflow, but is not included in this dataset release.

After teleoperation and before large-scale policy training:

• NVIDIA MimicGen was used to expand demonstrations

• ~10 human teleoperation rollouts were scaled to ~100 variations

• Randomized properties included:

  • Table color and textures
  • Lighting conditions
  • Scene appearance and material variations

The resulting domain-randomized datasets will be released at a later date, alongside trained policy weights and training configurations.

This work was conducted to deepen understanding of simulation fidelity and sim-to-real robustness beyond the base Lightwheel Leisaac tutorial.

---

Notes

• This dataset is intended for research and experimentation
• Raw format allows flexibility for custom pipelines
• File sizes are large due to full-fidelity simulation logging

---

Author

Priyanshu Sah (Hugging Face: xxxTEMPESTxxx)

---

License

License information will be added in a future update.

---

Future Updates

• Domain-randomized dataset release
• Policy weights and inference scripts
• Training configuration files
• Detailed evaluation results

Stay tuned for updates as the project progresses.