GeneralistAI – Research Preview of Dextrous Bimanual Robotic Manipulation
a year ago
- #AI
- #dexterity
- #robotics
- Generalist is developing general purpose robots using end-to-end AI models.
- Models are trained for dexterous sensorimotor policies across different robots and environments.
- Robots are controlled in real-time by neural networks processing sensor data at 100Hz.
- Tasks include fine motor control, spatial precision, and robustness to disturbances.
- Bi-manual coordination is used for complex actions like breaking apart Lego structures.
- High-frequency control enables behaviors like wiggling and adjusting in-flight grasps.
- Models generalize across different robot arms and new environments without prior data.
- Demonstrated tasks include sorting fasteners, folding boxes, and handling screws.
- Precision tasks like closing boxes require millimeter-level tolerances.
- The system shows potential for autonomous dexterity in real-world applications.