QT-Opt is a vision-based robotic manipulation dataset developed by researchers at Google and UC Berkeley. It focuses on closed-loop reinforcement learning for grasping tasks, containing over 580,000 real-world grasp attempts collected from a Sawyer robot. The dataset includes RGB camera observations, motor commands, and success/failure labels, enabling the training of deep neural networks for dynamic manipulation. QT-Opt's unique contribution lies in its ability to learn regrasping strategies, object probing, and dynamic responses to disturbances, achieving a 96% success rate on unseen objects. The dataset is released under the Open Data Commons Attribution License (ODC-BY), requiring attribution to the original authors while allowing modification and redistribution. It serves as a benchmark for vision-based RL research, demonstrating the potential of large-scale data collection for improving robotic dexterity and generalization.
Kuka robot picking objects in a bin.
| Field | Value |
|---|---|
| Action Space | EEF Position |
| Control Frequency | 10 |
| Depth Cams | 0 |
| Episodes | 200 |
| File Size | 0.59 GB |
| Gripper | Default |
| Has Camera Calibration | False |
| Has Proprioception | True |
| Has Suboptimal | False |
| Rgb Cams | 1 |
| Robot | Franka |
| Robot Morphology | Single Arm |
| Scene Type | Table Top |
| Wrist Cams | 0 |