DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset

arXiv: 2403.12945 | Venue: RSS 2024

Authors: Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, et al. (100+ contributors, 13 institutions)

Abstract

DROID (Distributed Robot Interaction Dataset) is a large-scale, geographically distributed robot manipulation dataset designed to advance policy generalization. Collected over 12 months across North America, Asia, and Europe, DROID comprises 76,000 human-teleoperated demonstration trajectories spanning 350 hours of interaction across 564 unique real-world scenes and 86 distinct manipulation tasks.

Dataset Scale

• Trajectories: 76,000
• Hours: 350
• Scenes: 564
• Tasks: 86
• Institutions: 13
• Robots: 18 Franka Panda 7-DoF arms
• Data collectors: 50

Hardware Setup

• Robot: Franka Panda 7-DoF arm
• Cameras: Two adjustable ZED 2 stereo cameras + wrist-mounted ZED Mini stereo camera
• Teleoperation device: Meta Oculus Quest 2 headset + controllers (6D pose control of arm + continuous gripper)
• Language annotations: 3 natural language annotations for 95% of successful episodes (75k episodes, added Dec 2024 on HuggingFace)

Key Design Decisions

• In-the-wild collection: Data gathered in diverse real-world environments (homes, offices, kitchens, labs) rather than controlled lab settings
• Distributed collection protocol: Standardized setup reproduced across 13 institutions globally
• Multi-camera: Three synchronized RGB camera streams + depth + calibration per episode
• Open release: Full dataset, policy learning code, and hardware reproduction guide released publicly

Performance Impact

Policies trained on DROID show higher performance and improved generalization compared to training on smaller, lab-only datasets. DROID is a key pretraining corpus for VLA models targeting real-world deployment.

Comparison Context

Dataset	Trajectories	Hours	Scenes	Tasks
BridgeData V2	60k	~240h	24 envs	82
DROID	76k	350h	564	86
RH20T	110k	—	—	147
Open-X Embodiment	1.4M	—	—	217+